Add Batch aa25d87e-b96c-4eda-9139-bb5ff0cc2732 data

.gitattributes

@@ -1254,3 +1254,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2025/kNN[[:space:]]Retrieval[[:space:]]for[[:space:]]Simple[[:space:]]and[[:space:]]Effective[[:space:]]Zero-Shot[[:space:]]Multi-speaker[[:space:]]Text-to-Speech/4c1bc120-0ab6-4523-9bb2-caf54e417ad4_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2024/A[[:space:]](More)[[:space:]]Realistic[[:space:]]Evaluation[[:space:]]Setup[[:space:]]for[[:space:]]Generalisation[[:space:]]of[[:space:]]Community[[:space:]]Models[[:space:]]on[[:space:]]Malicious[[:space:]]Content[[:space:]]Detection/ad9896f5-78a5-40e5-b19f-0ebb4789f513_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2024/A[[:space:]]Lightweight[[:space:]]Mixture-of-Experts[[:space:]]Neural[[:space:]]Machine[[:space:]]Translation[[:space:]]Model[[:space:]]with[[:space:]]Stage-wise[[:space:]]Training[[:space:]]Strategy/3333bd3e-943f-4661-ae40-014c8a021785_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/A[[:space:]]Measure[[:space:]]for[[:space:]]Transparent[[:space:]]Comparison[[:space:]]of[[:space:]]Linguistic[[:space:]]Diversity[[:space:]]in[[:space:]]Multilingual[[:space:]]NLP[[:space:]]Data[[:space:]]Sets/7dee6ef6-08dd-4565-9a55-2f0240bdda97_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/A[[:space:]]Novel[[:space:]]Paradigm[[:space:]]Boosting[[:space:]]Translation[[:space:]]Capabilities[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models/8eb804a5-0f10-4fdb-af4a-6a552f7a4c4e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/A[[:space:]]Novel[[:space:]]Two-step[[:space:]]Fine-tuning[[:space:]]Framework[[:space:]]for[[:space:]]Transfer[[:space:]]Learning[[:space:]]in[[:space:]]Low-Resource[[:space:]]Neural[[:space:]]Machine[[:space:]]Translation/07401a9f-9a22-4566-a6eb-41c799e41ff7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/A[[:space:]]Robust[[:space:]]Semantics-based[[:space:]]Watermark[[:space:]]for[[:space:]]Large[[:space:]]Language[[:space:]]Model[[:space:]]against[[:space:]]Paraphrasing/b1276659-b3bd-43b9-95dd-a7cb02ea87a5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/A[[:space:]]Study[[:space:]]on[[:space:]]Scaling[[:space:]]Up[[:space:]]Multilingual[[:space:]]News[[:space:]]Framing[[:space:]]Analysis/a4b2e349-9396-4656-a45a-c4d7d00460a3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/A[[:space:]]Systematic[[:space:]]Analysis[[:space:]]of[[:space:]]Subwords[[:space:]]and[[:space:]]Cross-Lingual[[:space:]]Transfer[[:space:]]in[[:space:]]Multilingual[[:space:]]Translation/d12825ff-73dc-4d7f-b285-a37a03d24322_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/A[[:space:]]Transformer[[:space:]]with[[:space:]]Stack[[:space:]]Attention/a415eb91-fbbd-49b8-a99a-1ef3cad21431_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/A[[:space:]]Tree-of-Thoughts[[:space:]]to[[:space:]]Broaden[[:space:]]Multi-step[[:space:]]Reasoning[[:space:]]across[[:space:]]Languages/cf769ade-bf6e-4081-af4e-c7800f171438_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/ADaPT_[[:space:]]As-Needed[[:space:]]Decomposition[[:space:]]and[[:space:]]Planning[[:space:]]with[[:space:]]Language[[:space:]]Models/9db614b1-10b0-4fc7-a80e-f25f41e0e9c5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/AGIEval_[[:space:]]A[[:space:]]Human-Centric[[:space:]]Benchmark[[:space:]]for[[:space:]]Evaluating[[:space:]]Foundation[[:space:]]Models/cc024075-7da6-4e18-80b6-15117eba05f2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/ATG_[[:space:]]Benchmarking[[:space:]]Automated[[:space:]]Theorem[[:space:]]Generation[[:space:]]for[[:space:]]Generative[[:space:]]Language[[:space:]]Models/50c4197c-f220-41e2-a316-82548d40ae14_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/AbsPyramid_[[:space:]]Benchmarking[[:space:]]the[[:space:]]Abstraction[[:space:]]Ability[[:space:]]of[[:space:]]Language[[:space:]]Models[[:space:]]with[[:space:]]a[[:space:]]Unified[[:space:]]Entailment[[:space:]]Graph/c07d0d6c-bc04-4fb5-ae02-b09747c9aab6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/AdaPT_[[:space:]]A[[:space:]]Set[[:space:]]of[[:space:]]Guidelines[[:space:]]for[[:space:]]Hyperbolic[[:space:]]Multimodal[[:space:]]Multilingual[[:space:]]NLP/238fa480-c7a8-4a4f-b80d-3e657c440dd4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/AdaRefiner_[[:space:]]Refining[[:space:]]Decisions[[:space:]]of[[:space:]]Language[[:space:]]Models[[:space:]]with[[:space:]]Adaptive[[:space:]]Feedback/fc84e3f8-62f5-4f67-9e4e-b40c92e70695_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Adapting[[:space:]]Fake[[:space:]]News[[:space:]]Detection[[:space:]]to[[:space:]]the[[:space:]]Era[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models/331f707f-051e-4194-a040-ac594764bd13_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Addressing[[:space:]]Both[[:space:]]Statistical[[:space:]]and[[:space:]]Causal[[:space:]]Gender[[:space:]]Fairness[[:space:]]in[[:space:]]NLP[[:space:]]Models/62035eb4-235e-43c3-ade0-3b324115f119_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Addressing[[:space:]]Healthcare-related[[:space:]]Racial[[:space:]]and[[:space:]]LGBTQ+[[:space:]]Biases[[:space:]]in[[:space:]]Pretrained[[:space:]]Language[[:space:]]Models/87e90e80-14ec-485d-a7c7-040084dbc277_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Adversarial[[:space:]]DPO_[[:space:]]Harnessing[[:space:]]Harmful[[:space:]]Data[[:space:]]for[[:space:]]Reducing[[:space:]]Toxicity[[:space:]]with[[:space:]]Minimal[[:space:]]Impact[[:space:]]on[[:space:]]Coherence[[:space:]]and[[:space:]]Evasiveness[[:space:]]in[[:space:]]Dialogue[[:space:]]Agents/58e96e8d-f02e-45ee-9796-d3cc69f1607f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Aligning[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]with[[:space:]]Recommendation[[:space:]]Knowledge/3f4152e5-2812-433e-bc14-3ecb6b14cc99_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/An[[:space:]]Effective[[:space:]]Automated[[:space:]]Speaking[[:space:]]Assessment[[:space:]]Approach[[:space:]]to[[:space:]]Mitigating[[:space:]]Data[[:space:]]Scarcity[[:space:]]and[[:space:]]Imbalanced[[:space:]]Distribution/e28af094-fb52-492a-9bc1-c317568cd6ff_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/An[[:space:]]End-to-End[[:space:]]Submodular[[:space:]]Framework[[:space:]]for[[:space:]]Data-Efficient[[:space:]]In-Context[[:space:]]Learning/c0a5d34d-6438-4e00-a34c-d108eadf53b8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Anonymity[[:space:]]at[[:space:]]Risk_[[:space:]]Assessing[[:space:]]Re-Identification[[:space:]]Capabilities[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]in[[:space:]]Court[[:space:]]Decisions/b40ab046-2f31-44f4-b831-76112146971f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Anti-LM[[:space:]]Decoding[[:space:]]for[[:space:]]Zero-shot[[:space:]]In-context[[:space:]]Machine[[:space:]]Translation/f1636993-024c-4d09-aad1-b7ad9762e1cb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Applications[[:space:]]of[[:space:]]BERT[[:space:]]Models[[:space:]]Towards[[:space:]]Automation[[:space:]]of[[:space:]]Clinical[[:space:]]Coding[[:space:]]in[[:space:]]Icelandic/7dc199e6-dbd0-46d3-b0e6-d88ae5767f5f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Asking[[:space:]]More[[:space:]]Informative[[:space:]]Questions[[:space:]]for[[:space:]]Grounded[[:space:]]Retrieval/e88afe77-44cc-4169-aad6-94f32e19deb9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Aspect-based[[:space:]]Sentiment[[:space:]]Analysis[[:space:]]with[[:space:]]Context[[:space:]]Denoising/d5810f74-09e6-4c47-934c-0beeba4c862c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Attention[[:space:]]Alignment[[:space:]]and[[:space:]]Flexible[[:space:]]Positional[[:space:]]Embeddings[[:space:]]Improve[[:space:]]Transformer[[:space:]]Length[[:space:]]Extrapolation/8ed8adf8-62dd-489c-ad39-c6864d107b10_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Automatic[[:space:]]Pair[[:space:]]Construction[[:space:]]for[[:space:]]Contrastive[[:space:]]Post-training/810c3832-4e4c-4ad4-964e-dbdda1174c63_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/BEAR_[[:space:]]A[[:space:]]Unified[[:space:]]Framework[[:space:]]for[[:space:]]Evaluating[[:space:]]Relational[[:space:]]Knowledge[[:space:]]in[[:space:]]Causal[[:space:]]and[[:space:]]Masked[[:space:]]Language[[:space:]]Models/aa29d969-de56-400b-a764-fc7e6b1ec79f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/BERTweet’s[[:space:]]TACO[[:space:]]Fiesta_[[:space:]]Contrasting[[:space:]]Flavors[[:space:]]On[[:space:]]The[[:space:]]Path[[:space:]]Of[[:space:]]Inference[[:space:]]And[[:space:]]Information-Driven[[:space:]]Argument[[:space:]]Mining[[:space:]]On[[:space:]]Twitter/c3374b87-597b-47f0-b03c-4eb3a8725b23_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Benchmarking[[:space:]]Generation[[:space:]]and[[:space:]]Evaluation[[:space:]]Capabilities[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]for[[:space:]]Instruction[[:space:]]Controllable[[:space:]]Summarization/410bbf8b-7b38-4f67-8a46-f62d156d2e1f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Best[[:space:]]of[[:space:]]Both[[:space:]]Worlds_[[:space:]]A[[:space:]]Pliable[[:space:]]and[[:space:]]Generalizable[[:space:]]Neuro-Symbolic[[:space:]]Approach[[:space:]]for[[:space:]]Relation[[:space:]]Classification/f2803076-e46c-4851-a3fa-67e000a36b95_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Beta-LR_[[:space:]]Interpretable[[:space:]]Logical[[:space:]]Reasoning[[:space:]]based[[:space:]]on[[:space:]]Beta[[:space:]]Distribution/8fc757ef-5ecc-49e5-bd1d-fd6e6aaa7a17_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Beyond[[:space:]]Read-Only_[[:space:]]Crafting[[:space:]]a[[:space:]]Comprehensive[[:space:]]Chinese[[:space:]]Text-to-SQL[[:space:]]Dataset[[:space:]]for[[:space:]]Database[[:space:]]Manipulation[[:space:]]and[[:space:]]Query/6db66d9c-0a02-4ac4-b234-ab89c90d2965_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Beyond[[:space:]]Surface[[:space:]]Similarity_[[:space:]]Detecting[[:space:]]Subtle[[:space:]]Semantic[[:space:]]Shifts[[:space:]]in[[:space:]]Financial[[:space:]]Narratives/5efd2f8b-4b62-4779-8668-4d37d4d240fe_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Bilateral[[:space:]]Masking[[:space:]]with[[:space:]]prompt[[:space:]]for[[:space:]]Knowledge[[:space:]]Graph[[:space:]]Completion/66d43eff-601d-4342-abc1-3231b613235d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Biomedical[[:space:]]Entity[[:space:]]Representation[[:space:]]with[[:space:]]Graph-Augmented[[:space:]]Multi-Objective[[:space:]]Transformer/9ff510b3-629c-4be2-b364-d5c3539b9cde_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/BotChat_[[:space:]]Evaluating[[:space:]]LLMs’[[:space:]]Capabilities[[:space:]]of[[:space:]]Having[[:space:]]Multi-Turn[[:space:]]Dialogues/bd76aead-d1f5-488f-97e8-f28a8a63c94f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Bring[[:space:]]Your[[:space:]]Own[[:space:]]KG_[[:space:]]Self-Supervised[[:space:]]Program[[:space:]]Synthesis[[:space:]]for[[:space:]]Zero-Shot[[:space:]]KGQA/95c71999-98f8-44a3-bb59-94ca8d0431cd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/C3LPGCN_Integrating[[:space:]]Contrastive[[:space:]]Learning[[:space:]]and[[:space:]]Cooperative[[:space:]]Learning[[:space:]]with[[:space:]]Prompt[[:space:]]into[[:space:]]Graph[[:space:]]Convolutional[[:space:]]Network[[:space:]]for[[:space:]]Aspect-based[[:space:]]Sentiment[[:space:]]Analysis/547fd9a7-a50e-4285-a9b4-0a007a91b8b9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/CARE_[[:space:]]Extracting[[:space:]]Experimental[[:space:]]Findings[[:space:]]From[[:space:]]Clinical[[:space:]]Literature/200eac9d-4ba1-4801-9f14-9f95c282fcdc_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/CLEAN–EVAL_[[:space:]]Clean[[:space:]]Evaluation[[:space:]]on[[:space:]]Contaminated[[:space:]]Large[[:space:]]Language[[:space:]]Models/6342f4bf-0929-4391-a2d9-86dd4ec781b1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/CLGSI_[[:space:]]A[[:space:]]Multimodal[[:space:]]Sentiment[[:space:]]Analysis[[:space:]]Framework[[:space:]]based[[:space:]]on[[:space:]]Contrastive[[:space:]]Learning[[:space:]]Guided[[:space:]]by[[:space:]]Sentiment[[:space:]]Intensity/4427be15-6040-4fe8-b74b-c333f65a94f0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/CM-TTS_[[:space:]]Enhancing[[:space:]]Real[[:space:]]Time[[:space:]]Text-to-Speech[[:space:]]Synthesis[[:space:]]Efficiency[[:space:]]through[[:space:]]Weighted[[:space:]]Samplers[[:space:]]and[[:space:]]Consistency[[:space:]]Models/ceb4a056-749e-4a6c-92b1-b167ec8b5f13_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/COMEM_[[:space:]]In-Context[[:space:]]Retrieval-Augmented[[:space:]]Mass-Editing[[:space:]]Memory[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Models/5cf99a4d-fef1-4522-900e-8b7bcec15b7e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/COMMIT_[[:space:]]Code-Mixing[[:space:]]English-Centric[[:space:]]Large[[:space:]]Language[[:space:]]Model[[:space:]]for[[:space:]]Multilingual[[:space:]]Instruction[[:space:]]Tuning/d0284585-1f25-4779-9927-6b8506f32ce9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Can[[:space:]]Public[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]Help[[:space:]]Private[[:space:]]Cross-device[[:space:]]Federated[[:space:]]Learning_/90d28b5b-b9ad-47ec-9d36-cdd31cd76eab_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Characterizing[[:space:]]Human[[:space:]]and[[:space:]]Zero-Shot[[:space:]]GPT-3.5[[:space:]]Object-Similarity[[:space:]]Judgments/fb4e2093-f58b-4cc3-967d-fceb9d8b9bb2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Chart-based[[:space:]]Reasoning_[[:space:]]Transferring[[:space:]]Capabilities[[:space:]]from[[:space:]]LLMs[[:space:]]to[[:space:]]VLMs/31393690-d0f7-4905-b99d-3996b8bda690_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Citation_[[:space:]]A[[:space:]]Key[[:space:]]to[[:space:]]Building[[:space:]]Responsible[[:space:]]and[[:space:]]Accountable[[:space:]]Large[[:space:]]Language[[:space:]]Models/b0037916-6f1e-4303-9add-33be19ce2d76_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/CoDa_[[:space:]]Constrained[[:space:]]Generation[[:space:]]based[[:space:]]Data[[:space:]]Augmentation[[:space:]]for[[:space:]]Low-Resource[[:space:]]NLP/57cba468-5d50-402f-9def-8487a1480ffa_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/CoMM_[[:space:]]Collaborative[[:space:]]Multi-Agent,[[:space:]]Multi-Reasoning-Path[[:space:]]Prompting[[:space:]]for[[:space:]]Complex[[:space:]]Problem[[:space:]]Solving/e8d29639-99fe-456f-8c2b-38fb1237cae8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/CodecLM_[[:space:]]Aligning[[:space:]]Language[[:space:]]Models[[:space:]]with[[:space:]]Tailored[[:space:]]Synthetic[[:space:]]Data/d558e8e0-cba3-4b35-b44a-38e08524730d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Cognitive[[:space:]]Overload_[[:space:]]Jailbreaking[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]with[[:space:]]Overloaded[[:space:]]Logical[[:space:]]Thinking/79c8f2df-d921-4b49-9eba-62bda3b86dd3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Comparing[[:space:]]Two[[:space:]]Model[[:space:]]Designs[[:space:]]for[[:space:]]Clinical[[:space:]]Note[[:space:]]Generation;[[:space:]]Is[[:space:]]an[[:space:]]LLM[[:space:]]a[[:space:]]Useful[[:space:]]Evaluator[[:space:]]of[[:space:]]Consistency_/2243cb5e-46cd-403d-b0c8-5c105eecbe7c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Compensate[[:space:]]Quantization[[:space:]]Errors_[[:space:]]Make[[:space:]]Weights[[:space:]]Hierarchical[[:space:]]to[[:space:]]Compensate[[:space:]]Each[[:space:]]Other/a93433db-0136-4997-a3f4-b3c917180f82_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Composite[[:space:]]Backdoor[[:space:]]Attacks[[:space:]]Against[[:space:]]Large[[:space:]]Language[[:space:]]Models/e0b2b567-e32e-4bba-b7db-d610c1358e3a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Conformal[[:space:]]Intent[[:space:]]Classification[[:space:]]and[[:space:]]Clarification[[:space:]]for[[:space:]]Fast[[:space:]]and[[:space:]]Accurate[[:space:]]Intent[[:space:]]Recognition/dbb317c0-cdaf-4bb1-9919-afb55efe3cd2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Connecting[[:space:]]the[[:space:]]Dots_[[:space:]]Inferring[[:space:]]Patent[[:space:]]Phrase[[:space:]]Similarity[[:space:]]with[[:space:]]Retrieved[[:space:]]Phrase[[:space:]]Graphs/28e48b91-36fc-443a-963a-cf3c4894b2f0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Content-Specific[[:space:]]Humorous[[:space:]]Image[[:space:]]Captioning[[:space:]]Using[[:space:]]Incongruity[[:space:]]Resolution[[:space:]]Chain-of-Thought/c41e0981-e78c-4b2c-85c9-80eb39a340f9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Context[[:space:]]Does[[:space:]]Matter_[[:space:]]Implications[[:space:]]for[[:space:]]Crowdsourced[[:space:]]Evaluation[[:space:]]Labels[[:space:]]in[[:space:]]Task-Oriented[[:space:]]Dialogue[[:space:]]Systems/7457564d-3558-4f72-8485-4163d0d57888_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Contrastive[[:space:]]Learning[[:space:]]as[[:space:]]a[[:space:]]Polarizer_[[:space:]]Mitigating[[:space:]]Gender[[:space:]]Bias[[:space:]]by[[:space:]]Fair[[:space:]]and[[:space:]]Biased[[:space:]]sentences/bbce1d22-4ed9-4b19-afeb-233233896a8f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Contrastive[[:space:]]Preference[[:space:]]Learning[[:space:]]for[[:space:]]Neural[[:space:]]Machine[[:space:]]Translation/141e89b4-8423-4cdb-aa1d-9c2807e95392_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Crafting[[:space:]]In-context[[:space:]]Examples[[:space:]]according[[:space:]]to[[:space:]]LMs’[[:space:]]Parametric[[:space:]]Knowledge/9859eb22-4387-4372-a7cc-099a5a84595d_origin.pdf filter=lfs diff=lfs merge=lfs -text

2024/A Measure for Transparent Comparison of Linguistic Diversity in Multilingual NLP Data Sets/7dee6ef6-08dd-4565-9a55-2f0240bdda97_content_list.json

@@ -0,0 +1,2002 @@
+[
+    {
+        "type": "text",
+        "text": "A Measure for Transparent Comparison of Linguistic Diversity in Multilingual NLP Data Sets",
+        "text_level": 1,
+        "bbox": [
+            157,
+            79,
+            840,
+            118
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Tanja Samardžić",
+        "bbox": [
+            174,
+            126,
+            326,
+            142
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Language and Space Lab",
+        "bbox": [
+            146,
+            143,
+            352,
+            158
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "University of Zurich",
+        "bbox": [
+            166,
+            160,
+            334,
+            175
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "tanja.samardzic@uzh.ch",
+        "bbox": [
+            139,
+            175,
+            361,
+            191
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Ximena Gutierrez",
+        "bbox": [
+            418,
+            126,
+            579,
+            140
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "CEIICH",
+        "bbox": [
+            463,
+            143,
+            532,
+            156
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Universidad Nacional",
+        "bbox": [
+            410,
+            160,
+            588,
+            173
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Autónoma de México",
+        "bbox": [
+            410,
+            175,
+            588,
+            190
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Christian Bentz",
+        "bbox": [
+            678,
+            126,
+            820,
+            141
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Dept. of General Linguistics",
+        "bbox": [
+            633,
+            143,
+            865,
+            158
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Eberhard-Karls-Universität",
+        "bbox": [
+            636,
+            160,
+            860,
+            174
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Tübingen",
+        "bbox": [
+            710,
+            175,
+            789,
+            192
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Steven Moran",
+        "bbox": [
+            268,
+            204,
+            393,
+            219
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Laboratory of Language Evolution\nUniversity of Neuchâtel",
+        "bbox": [
+            189,
+            221,
+            472,
+            253
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Olga Pelloni",
+        "bbox": [
+            611,
+            204,
+            721,
+            219
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Telepathy Labs GmbH",
+        "bbox": [
+            573,
+            221,
+            759,
+            237
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Zurich, Switzerland",
+        "bbox": [
+            584,
+            237,
+            747,
+            252
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            261,
+            339,
+            275
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Typologically diverse benchmarks are increasingly created to track the progress achieved in multilingual NLP. Linguistic diversity of these data sets is typically measured as the number of languages or language families included in the sample, but such measures do not consider structural properties of the included languages. In this paper, we propose assessing linguistic diversity of a data set against a reference language sample as a means of maximising linguistic diversity in the long run. We represent languages as sets of features and apply a version of the Jaccard index  $(J_{mm})$  suitable for comparing sets of measures. In addition to the features extracted from typological data bases, we propose an automatic text-based measure, which can be used as a means of overcoming the well-known problem of data sparsity in manually collected features. Our diversity score is interpretable in terms of linguistic features and can identify the types of languages that are not represented in a data set. Using our method, we analyse a range of popular multilingual data sets (UD, Bible100, mBERT, XTREME, XGLUE, XNLI, XCOPA, TyDiQA, XQuAD). In addition to ranking these data sets, we find, for example, that (poly)synthetic languages are missing in almost all of them.",
+        "bbox": [
+            141,
+            290,
+            460,
+            703
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            717,
+            258,
+            732
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Data sets for training and testing NLP models are increasingly multilingual and aimed at broad linguistic coverage. These data sets are often claimed to represent a typologically diverse sample, including low-resource and endangered languages.",
+        "bbox": [
+            112,
+            744,
+            489,
+            824
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Linguistic diversity is typically described as the number of languages included in the data set, yet less often as the number of language families to which these languages belong. Both counts indicate a level of linguistic diversity: the more languages and families, the more diversity. But how",
+        "bbox": [
+            112,
+            825,
+            489,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "do we know that included languages are indeed different? How can we define a desired or optimal diversity to set as a goal when composing multilingual data sets? These questions need to be addressed if our goal is to know how NLP technology generalises across diverse languages, without testing it on each single language (even if we had the necessary data for all languages).",
+        "bbox": [
+            507,
+            261,
+            884,
+            390
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The aim of this paper is to initiate and facilitate comparisons between multilingual NLP data sets with respect to a linguistic diversity reference. For this, we propose a measure of linguistic diversity and a method of comparison that identifies what kinds of linguistic features are missing. As an initial reference, we rely on a predefined sample of languages — the 100-language-sample (100L) selected by the Word Atlas of Language Structures (WALS; Comrie et al. (2013)) to represent geographic and phylogenetic diversity. As a comparison method, we formulate a version of the Jaccard index suitable for comparing measures. This measure allows us to quantify the distance between the observed and the reference diversity in terms of linguistic features, showing not only how diverse language samples are but also what kinds of linguistic phenomena are not represented in a given sample. To facilitate automatic extraction of linguistic features needed for assessing linguistic diversity, we complement the information from linguistic data bases with relevant text statistics.",
+        "bbox": [
+            507,
+            406,
+            884,
+            760
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Our proposals are intended to help researchers make informed choices when designing a multilingual data set. Representing a wider spectrum of linguistic diversity is not only a way to improve the cross-linguistic generalisation of NLP technology, but also a way to deal with biases against low-resource languages, which are harder to represent and thus more likely to be left behind (Joshi et al., 2020).",
+        "bbox": [
+            507,
+            777,
+            884,
+            920
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "3367",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: NAACL 2024, pages 3367-3382",
+        "bbox": [
+            218,
+            945,
+            778,
+            958
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "June 16-21, 2024 ©2024 Association for Computational Linguistics",
+        "bbox": [
+            292,
+            959,
+            705,
+            971
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/eb7d8c18bf9e049512d0759f01f3135b27f9ce22c7df6c433a234891f914192e.jpg",
+        "image_caption": [
+            "Figure 1: Languages in the WALS 100L sample with their endangerment status."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            84,
+            880,
+            350
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Background and Related Work",
+        "text_level": 1,
+        "bbox": [
+            112,
+            409,
+            418,
+            425
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Evaluating the linguistic diversity of data sets relies on comparable descriptions of languages. For instance, the (approximate) number of speakers is an attribute whose value can be found and compared for all registered languages. This attribute, however, does not describe the structure of languages. An example of a structural attribute would be the presence or the absence of adjectives in a language. To establish the value of this attribute for any language, we need a universal definition of what an adjective is. It turns out that such universal definitions are hard to formulate in a principled way (Haspelmath, 2007), which makes it hard to define objective measures of how similar or dissimilar any two languages are.",
+        "bbox": [
+            112,
+            437,
+            489,
+            678
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The most widely accepted method for comparing languages relies on genealogical classification: given a phylogenetic tree, we consider languages located in the same region of the tree to be similar. This method currently prevails in NLP (cf. the work discussed in Section 6). Typically, we regard languages that belong to the same family to be similar. To know which language belongs to which family, we turn to popular authorities such as WALS (Dryer and Haspelmath, 2013) or Glottolog (Hammarström et al., 2018). However, language families can be too broad for a meaningful comparison as they include typologically very different languages. For instance, English and Armenian belong to the same family, Indo-European, but are",
+        "bbox": [
+            112,
+            680,
+            489,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "vastly different in terms of their phoneme inventories, morphology, and word order.",
+        "bbox": [
+            507,
+            409,
+            884,
+            439
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Another possibility to compare languages, starting to be used in NLP only recently, is to rely on grammatical features available in the WALS data base. This is a comprehensive source of information about linguistic structures but still rather sparsely populated; feature values are often known for only a few languages. $^{1}$  Together with other typological data bases, WALS is included in URIEL, an aggregated and standardised source of language features for various NLP uses. Ponti et al. (2020) propose a diversity score using the features from URIEL (Littell et al., 2017). The score is called typology index and it is calculated as the entropy of feature values (averaged per data set). $^{2}$  In other NLP work, grammatical features (usually termed typological) are used for other purposes, such as predicting the features (Ponti et al., 2019) rather than using them for language sampling in creating multilingual data sets. Moran (2016) use WALS and AUTOTYP features (Stoll and Bickel, 2013) to compose a sample of 10 maximally diverse languages for a corpus-based study of language acquisition.",
+        "bbox": [
+            507,
+            442,
+            884,
+            810
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Finally, languages can be described using features derived from various text statistics, but such features are not commonly used for language sam",
+        "bbox": [
+            507,
+            812,
+            884,
+            860
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>1</sup>An alternative typological data base is AUTOTYP (Bickel et al., 2017), with a different design but similar coverage.",
+        "bbox": [
+            507,
+            870,
+            882,
+            895
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_footnote",
+        "text": "2They propose two more scores, family and geography, which do not make use of grammatical features.",
+        "bbox": [
+            507,
+            895,
+            882,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "3368",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "pling. Type-token ratio (TTR) or unigram entropy of a text have been shown to correlate with grammar-based morphological complexity measures (Kettunen, 2014; Bentz et al., 2016). Many other methods have been proposed for assessing linguistic complexity using text statistics (see, for instance, Berdicevskis et al. (2018)). All of these measures can, in principle, be used for describing and comparing languages although such comparisons might seem counter-intuitive and hard to interpret in terms of genealogical classification. On the other hand, these features might complement usual descriptions of languages while being more directly relevant to text processing and NLP.",
+        "bbox": [
+            112,
+            84,
+            489,
+            309
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Transfer learning created a new need for nuanced languages comparison for NLP. While models can now be transferred across languages with zero-shot or few-shot learning (Pires et al., 2019), the success of the transfer might depend on the differences between languages. Lin et al. (2019) propose a range of measures that can be used in order to choose the best transfer language, which they divide into data-dependent (data size, token overlap, TTR) and data independent (various distance measures extracted from the URIEL data base). Lauscher et al. (2020) study how well different similarity scores predict the success of the transfer and they find that language family is, in fact, the one that is least helpful in all the tasks considered (with mBERT and XLM-R). Various criteria for assessing language similarity remain an open research area in NLP (Turc et al., 2021; Pelloni et al., 2022; Samardžić et al., 2022; de Vries et al., 2022). Our proposal for assessing linguist diversity is relevant to these efforts too, as its key component is language comparison at the level of features extracted from both typological data bases and text samples.",
+        "bbox": [
+            115,
+            311,
+            489,
+            681
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "More generally, our work is intended to contribute to several wide-scope initiatives for improving the quality of data management in multilingual NLP (Bender and Friedman, 2018; Kreutzer et al., 2021; Lhoest et al., 2021) by focusing specifically on diversity assessments and data-independent scores for language comparison.",
+        "bbox": [
+            112,
+            683,
+            489,
+            796
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Comparing Data Sets with Jaccard Similarity",
+        "text_level": 1,
+        "bbox": [
+            112,
+            810,
+            450,
+            845
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Our goal is to estimate the linguistic diversity of a data set with respect to some reference. Our score is thus a comparison between two data sets. More precisely, we compare scaled distributions",
+        "bbox": [
+            112,
+            856,
+            489,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/5c44663fd5263138d2cc361837a03ac53382f7ae4db8776ed7f1b2ece09141bf.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            556,
+            84,
+            685,
+            149
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/485964166e77c9c029e9e78008cce107d4168c89fe49bb96e740f4df28fa3e8f.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            705,
+            85,
+            833,
+            149
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/072f5981cd83cb0581fea1cbe784ee622b81eb71488e4057aa66fd414821e838.jpg",
+        "image_caption": [
+            "Figure 2: A toy example of comparing sets of measures with the minmax version of the Jaccard index."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            626,
+            152,
+            752,
+            225
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "of the values of a numerical attribute as shown in Figure 2. The upper part of the figure shows (constructed) examples of two data sets (A and B), which we compare assuming that A is the data set whose diversity we want to assess and B is the reference. The values of the numerical attribute (one measurement per language) are on the x-axis and the numbers of languages are on the y-axis. Each bar in the figures represents the number of languages in the given data set with the numerical value in the given range (bin). For instance, the first bar in the upper left plot shows that the first sample (A) has 30 languages, with the values of their numerical attributes falling between 1 and 2. The other sample (B) has no languages in this bin.",
+        "bbox": [
+            505,
+            293,
+            884,
+            533
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The width of the bins is arbitrary, but it does impact the score. Narrower bins capture more differences between two distributions than wider bins. By setting the width of the bins, we thus control the resolution at which we want to compare two data sets. In our example, the width is the distance between integers, but one can define other values (as long as the bins are of the same width).",
+        "bbox": [
+            507,
+            533,
+            882,
+            662
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Since the data sets that we compare contain different numbers of languages, the values on the y-axis (counts of languages) are normalised in order to neutralise the effect of the size of the samples and focus rather on the diversity. We multiply all counts in the smaller set with the scalar  $c$ :",
+        "bbox": [
+            507,
+            663,
+            882,
+            758
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nc = \\frac {\\operatorname {m a x} (| A | , | B |)}{\\operatorname {m i n} (| A | , | B |)} \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            618,
+            771,
+            880,
+            806
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this way, we increase the counts in the smaller set proportionally to obtain the same number of data points in both distributions and comparable numbers in each bin.",
+        "bbox": [
+            507,
+            812,
+            882,
+            875
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_footnote",
+        "text": "Another way to normalise the counts would be to divide them by the size of the set, but we chose the first option in order to preserve the notion of number of languages, which is",
+        "bbox": [
+            507,
+            883,
+            882,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "3369",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Once we have represented our two sets in this way, we compare them using a generalised version of Jaccard similarity. This score shows how much the two distributions overlap. The original Jaccard index (Jaccard, 1912) compares two sets, but its generalised versions are suitable for comparing sets of measurements. Thus, we use the minmax version of the score  $(J_{mm})$ , initially proposed by Tanimoto (1958) for comparing vectors of binary values and then generalised to weight vectors by Grefenstette (1994). In our version, we compare two data sets as two vectors of weights: each bin is one dimension in the vectors and the number of languages in that bin is its weight.",
+        "bbox": [
+            110,
+            84,
+            487,
+            309
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Intuitively, the score is the ratio between the intersection and the union of the two distributions (shown in the bottom part of Figure 2). Formally, we first map all the languages in all data sets to real numbers  $m:\\mathbb{L}\\mapsto \\mathbb{R}$ , so that  $\\{Y = m(x):x\\in X\\} = \\{(x_i,y_i)\\}$ , where  $x$  is a language in a data set,  $y$  is its corresponding measurement  $(y\\in \\mathbb{R})$  and the range of the index  $i1\\dots |X|$  is the set of languages included in a data set. We then group the measurements into bins by applying a given threshold:  $\\{Z = t(y):y\\in Y\\} = \\{(y_i,z_j)\\}$ , where  $z$  is the bin to which the measurement is assigned, the range of  $i$  is  $1\\dots |X|$  and the range of  $j$  is  $1\\dots |Z|$ .",
+        "bbox": [
+            112,
+            312,
+            489,
+            537
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "With this formalisation, we define the Jaccard minmax similarity of two data sets,  $J_{mm}(A,B)$ , as a similarity between two vectors of weights:",
+        "bbox": [
+            112,
+            541,
+            487,
+            590
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nJ _ {m m} (\\mathbf {a}, \\mathbf {b}) = \\frac {\\sum_ {j = 1} ^ {| Z |} \\operatorname* {m i n} \\left(a _ {j} , b _ {j}\\right)}{\\sum_ {j = 1} ^ {| Z |} \\operatorname* {m a x} \\left(a _ {j} , b _ {j}\\right)} \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            173,
+            612,
+            485,
+            659
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The sum in the numerator represents the intersection and the sum in the denominator the union of the two sets of measurements. The weights  $a$  and  $b$  represent the number of measurements in the bin  $j$ . The values of  $J_{mm}$  fall in the range [0, 1], with higher values indicating more similarity between A and B, and, indirectly, better coverage of linguistic diversity in A.",
+        "bbox": [
+            112,
+            677,
+            487,
+            806
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "What is especially interesting about using  $J_{mm}$  as a diversity score is its transparency in terms of individual measurements: we can visualise and interpret where exactly a data set departs from the reference.",
+        "bbox": [
+            112,
+            809,
+            487,
+            888
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4 Language Features",
+        "text_level": 1,
+        "bbox": [
+            507,
+            83,
+            709,
+            99
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We now turn to the question of how to define and take measures (the values on the x-axis in Figure 2) that can be used for calculating Jaccard minmax similarity between sets of languages. We use two kinds of descriptions.",
+        "bbox": [
+            507,
+            112,
+            882,
+            193
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1 Grammar Features",
+        "text_level": 1,
+        "bbox": [
+            507,
+            211,
+            707,
+            225
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Typological data bases are currently the principal source of information about the properties of languages, but NLP researchers are faced with many obstacles when using this information. The popular software package lang2vec associated with the URIEL data base (Littell et al., 2017) alleviates some of the obstacles. First, the package solves the problem of incompatible feature values across different sources by mapping the data from several original data bases to binary features. Second, the problem of sparsity of feature values is solved by imputing the missing values: instead of a missing feature value in a language, the package returns the observed value for the same feature in the closest language. In this way, features become available for all queried languages, which is necessary for estimating language diversity, but a large proportion (roughly  $40\\%$ ) of the returned features are imputed.",
+        "bbox": [
+            507,
+            234,
+            884,
+            524
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "While lang2vec facilitates retrieving typological features, its use for describing languages is limited due to remaining obstacles that are hard to solve. First, it does not contain any morphological features, which are especially relevant to NLP due to the known difficulties with that morphologically rich languages (Tsarfaty et al., 2013). The second unsolved problem is the fact that typological features are hard to add for languages for which they are not already available. Adding new features requires human expertise in many languages.",
+        "bbox": [
+            507,
+            526,
+            882,
+            703
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.2 Text Features",
+        "text_level": 1,
+        "bbox": [
+            507,
+            720,
+            662,
+            734
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "As a complement to commonly used features from lang2vec, we make use of linguistically relevant text statistics. In this study, we focus on the mean word length as an approximation of aggregated morphological features, but other text-based features might be envisioned in future work. Our choice to start with word length relies on the observation that longer words can be expected in languages with rich morphology (large morphological paradigms, productive derivation), while shorter words are found in languages with less morphol-",
+        "bbox": [
+            507,
+            744,
+            882,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "helpful for the subsequent explanations.",
+        "bbox": [
+            112,
+            906,
+            359,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "3370",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "ogy. $^{4}$  As an empirical confirmation of the relationship between the word length and morphology, we perform a correlation test between the mean word length and morphological complexity calculated over morphological features (see Section 5 for the methods and Section 7 for the discussion).",
+        "bbox": [
+            112,
+            84,
+            487,
+            179
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Text features are especially interesting in the context of NLP because they can be calculated automatically and applied to any language in which there are any texts to process. An important advantage of word length over other text statistics in this regard is that it manifests itself in very small samples of text and remains stable across different sizes. A sample of contiguous text of only 500 tokens gives us already a very good estimation of the overall mean word length. This can be seen in Figure 4 in the Appendix A, which shows the values of the mean word length on random samples of the length 500, 2000 and 10000 tokens in 87 languages. A correlation score (also in the Appendix A) shows that languages are almost identically ranked with all the sample sizes.",
+        "bbox": [
+            115,
+            181,
+            489,
+            439
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.3 Maximising Linguistic Diversity",
+        "text_level": 1,
+        "bbox": [
+            112,
+            454,
+            413,
+            469
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The editors of the WALS data base have selected two samples of languages (100 and 200 sample) as a means of guidance in the collective effort to create linguistic descriptions on a wide scale. These samples maximise genealogical (language family) and areal (geographic) diversity. Completing their descriptions is expected to minimise a potential bias regarding the relative frequency of different types of linguistic features included in the data base (Comrie et al., 2013). Figure 1 shows the locations of the languages in the 100 sample and their endangerment status according to UNESCO.",
+        "bbox": [
+            112,
+            476,
+            489,
+            668
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Recently, text samples have been collected for most of the 100 languages in the TeDDi data set (Moran et al., 2022). These text data are sampled from online resources, e.g., Project Gutenberg, Open Subtitles (Lison and Tiedemann, 2016), The Parallel Bible Corpus (Mayer and Cysouw, 2014), the Universal Declaration of Human Rights, but also from grammars and other language documentation sources. For languages not present in online resources, the texts were manually transcribed.",
+        "bbox": [
+            112,
+            671,
+            489,
+            832
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We take these two resources as the current reference that maximises linguistic diversity in terms of grammar features (WALS) and text features (TeDDi). We compare NLP data sets with these references, but our method can be applied to compare any given pair of data sets including potentially better references in the future.",
+        "bbox": [
+            507,
+            84,
+            884,
+            197
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5 Data and Methods",
+        "text_level": 1,
+        "bbox": [
+            507,
+            208,
+            705,
+            224
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We calculate the Jaccard minmax diversity score  $(\\mathrm{J}_{mm})$  for a number of popular multilingual data sets in comparison to the TeDDi sample. Without attempting to provide an exhaustive evaluation, we review data sets that satisfy the following criteria: multilingual (containing ten or more languages), relatively widely used and recently released or updated. The list is given in Table 1 and discussed in more detail in Section 6. For reference, we compare our  $\\mathrm{J}_{mm}$  score to the typological index (TI) previously proposed as a linguistic diversity measure by Ponti et al. (2020) (see Section 2).",
+        "bbox": [
+            507,
+            233,
+            884,
+            426
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Descriptions of the data sets often do not include all the information that was needed for our comparison. In particular, the number of language families is often not stated. To add this information, we extracted language names from the data files, converted these names into ISO 639-3 codes manually, and then retrieved the corresponding families from the Glottolog data base (top level family). Note that the conversion to ISO 639-3 codes led to some changes in the number of languages, compared to those cited in the data descriptions. For instance, the mBERT training data has only 97 distinct languages, not 104 as mentioned in the original description.",
+        "bbox": [
+            507,
+            426,
+            885,
+            652
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5.1 Methods for Text Features",
+        "text_level": 1,
+        "bbox": [
+            507,
+            663,
+            764,
+            678
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We define words to be sequences of Unicode characters, delimited by spaces or other language-specific word delimiters, as defined by common multilingual tokenisers. We tokenise all the collected samples into word-level tokens using the Python library Polyglot (Al-Rfou, 2015). If a resulting token does not contain any alphanumeric characters, we discard it as punctuation. All the remaining tokens are further segmented into characters using the Python library segments (Moran and Cysouw, 2018). We split words into sequences of",
+        "bbox": [
+            507,
+            683,
+            884,
+            860
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "4We give a more specific definition of the notion of a word as part of the methods in Section 5.",
+        "bbox": [
+            112,
+            844,
+            487,
+            869
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>5</sup>https://github.com/MorphDiv/TeDDi_sample/master",
+        "bbox": [
+            139,
+            870,
+            453,
+            894
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{6}$ https://www.gutenberg.org/",
+        "bbox": [
+            137,
+            894,
+            342,
+            906
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>7</sup>http://unicode.org/udhr/",
+        "bbox": [
+            137,
+            906,
+            327,
+            919
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "The code for reproducing the calculations can be found at https://github.com/MorphDiv/jmm_diversity/.",
+        "bbox": [
+            507,
+            869,
+            882,
+            894
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{9}$ https://polyglot.readthedocs.io",
+        "bbox": [
+            532,
+            894,
+            774,
+            906
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "10https://github.com/cldf/segments",
+        "bbox": [
+            532,
+            906,
+            781,
+            919
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "3371",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "characters and take their length as word length. $^{11}$  We apply this same definition to all scripts, but we discuss below potential adjustments in the case of (partially) logographic scripts.",
+        "bbox": [
+            112,
+            84,
+            487,
+            148
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Since the mean word length can be calculated on small samples, we take a single random sample for each language in a data set that we consider. To do this, we select a random position in the data set and extract contiguous text of the length up to 10K tokens starting from the random position. In case a data set does not contain such long texts (or sequences of paragraphs), we take smaller samples. The smallest samples are 200-300 tokens long.",
+        "bbox": [
+            112,
+            149,
+            487,
+            292
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The output of these text processing steps is a set of real numbers, each number representing a language in a data set. To turn these numbers into discrete features, we group them into bins of equal size. We set the bin width to 1.[12]",
+        "bbox": [
+            112,
+            294,
+            487,
+            372
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Mean word length vs. WALS features Following Bentz et al. (2016), we calculate a complexity score  $(C_{WALS})$  for each language using the set of 26 features that are relevant to describing morphology. This score is obtained by: 1) transforming the range of values each feature can take so that bigger values reflect the increasing use of morphology; 2) normalizing and averaging the resulting feature values per language. The list of features and transformations is given in Table 4 in Appendix B.  $C_{WALS}$  ranges from 0 to 1, where values closer to one indicate that the language encodes more morphosyntactic distinctions, making its morphology richer. All the values of the mean word length and morphological complexity for 29 diverse languages (the subset of TeDDi languages for which the 26 WALS features are known) are shown in Table 3 in Appendix B. We observe a strong correlation  $(\\rho = 0.69)$ , which means that the variables quantify very similar phenomena and that the mean word length is a reasonable approximation of morphological types of languages. We return to this point in Section 7.",
+        "bbox": [
+            112,
+            376,
+            489,
+            746
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Adjustments for logographic scripts Words in languages with logographic scripts tend to be shorter due to the fact that a single symbol corresponds to several alphabetic symbols (Sproat and Gutkin, 2021). For instance, in Mandarin Chinese, types such as the de (possessive particle), the le (aspect particle), is shi (copular verb 'is'), our wǒ-",
+        "bbox": [
+            112,
+            750,
+            489,
+            863
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "men (pronoun 'us') are assigned lengths (1, 1, 1, 2) respectively when measured in UTF-8 characters in the original script. When transliterated into Pinyin, the corresponding lengths are (2, 2, 3, 5). Hence, compared to Pinyin, the lengths are somewhat underestimated. It might seem more appropriate to convert the logographic scripts into their romanised counterparts to achieve cross-linguistic comparability. We opt for leaving such scripts without conversion, because we consider this phenomenon part of the diversity that we want to capture. Additional motivation for our choice is the fact that NLP systems have to deal with text as it is regardless of the mapping between written characters and sounds.",
+        "bbox": [
+            507,
+            84,
+            884,
+            309
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Three languages in our data samples, Chinese, Japanese and Korean, are affected by this issue to a varied degree. In these cases, we scale the observed word length proportionally to the difference between the Chinese original script and Pinyin so that the scaled length is comparable to alphabetic scripts. Table 5 in Appendix C shows revised diversity scores after the adjustments.",
+        "bbox": [
+            507,
+            310,
+            884,
+            439
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.2 Linguistic Diversity Scores",
+        "text_level": 1,
+        "bbox": [
+            507,
+            451,
+            764,
+            468
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "With the grammar features extracted from URIEL, we calculate syntactic diversity according to both TI and  $\\mathbf{J}_{mm}$ .",
+        "bbox": [
+            507,
+            474,
+            882,
+            523
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Syntax Typological Index  $(\\mathbf{TI}_{syn})$  Following the formulation by Ponti et al. (2020), we calculate the typological index for each data set. In this context, a language is characterized by 103 syntactic features with binary values<sup>13</sup>. For each feature, Shannon entropy is estimated using the distribution of feature values in a data set. The feature-specific entropy values are averaged over the full set of features to obtain a TI score ranging from 0 to 1. The TI values closer to 1 indicate more diversity.",
+        "bbox": [
+            507,
+            527,
+            882,
+            688
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Syntax Jaccard (J_mm_syn) We apply Jaccard similarity for comparing each data set against the TeDDi sample. Here the measures are the counts of the observed values of the same 103 syntactic feature available in 1ang2vec. This means that the items on the x-axis in Figure 2 are the 103 values, while the y-axis represents the number of times each feature value was observed in a data set. Since these feature values are binary, the width of the bin is not arbitrary in this case; it is determined by the values. Conceptually, grouping several features",
+        "bbox": [
+            507,
+            694,
+            882,
+            871
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>11</sup>We use the units defined by the Unicode Standard as \"user-perceived characters\" (NFC).",
+        "bbox": [
+            112,
+            869,
+            487,
+            895
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_footnote",
+        "text": "12 In addition to this, we also tried smaller bin sizes. We do not report the latter results, but the main trends did not change.",
+        "bbox": [
+            112,
+            895,
+            487,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>13</sup>We use the syntax_knn features available in lang2vec, which includes predicted values for those languages whose features are not available",
+        "bbox": [
+            507,
+            882,
+            882,
+            919
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "3372",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/390ee998200278f2f6803e6fc44a26928bdb0a275e417ce523fe2db1b556a245.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Name and main references</td><td>N(L)</td><td>N(F)</td><td>TI_syn</td><td>Jmm_syn</td><td>TI_morph</td><td>Jmm_morph</td></tr><tr><td>Universal Dependencies (UD)</td><td>106*</td><td>20*</td><td>0.567</td><td>0.736</td><td>0.349</td><td>0.650</td></tr><tr><td>Bible 100</td><td>103*</td><td>30*</td><td>0.649</td><td>0.811</td><td>0.311</td><td>0.534</td></tr><tr><td>mBERT</td><td>97*</td><td>15*</td><td>0.559</td><td>0.710</td><td>0.323</td><td>0.603</td></tr><tr><td>XTREME</td><td>40</td><td>14</td><td>0.612</td><td>0.775</td><td>0.311</td><td>0.457</td></tr><tr><td>XGLUE</td><td>19</td><td>7*</td><td>0.517</td><td>0.674</td><td>0.307</td><td>0.504</td></tr><tr><td>XNLI</td><td>15</td><td>7*</td><td>0.557</td><td>0.711</td><td>0.339</td><td>0.598</td></tr><tr><td>XCOPA</td><td>11</td><td>11</td><td>0.586</td><td>0.737</td><td>0.361</td><td>0.608</td></tr><tr><td>TyDiQA</td><td>11</td><td>10</td><td>0.626</td><td>0.751</td><td>0.343</td><td>0.525</td></tr><tr><td>XQuAD</td><td>12*</td><td>6*</td><td>0.523</td><td>0.680</td><td>0.341</td><td>0.588</td></tr><tr><td>TeDDi</td><td>89</td><td>51</td><td>0.706</td><td>-</td><td>0.369</td><td>-</td></tr></table>",
+        "bbox": [
+            122,
+            80,
+            877,
+            262
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Diversity of multilingual NLP data sets. The numbers in the second and the third column marked with an asterisk are added or modified by us. The numbers without an asterisk are reported in the respective publications. N(L): the number of languages in the data set. N(F): the number of families to which the languages belong. TI: typology index Ponti et al. (2020).  $J_{mm}$ : Jaccard minmax similarity (this paper).",
+        "bbox": [
+            112,
+            271,
+            884,
+            330
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "into a single one would correspond to increasing the bin width, but it is not clear at the moment how the features could be grouped. We thus work with the original set without any changes.",
+        "bbox": [
+            112,
+            354,
+            487,
+            419
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "With text features (mean word length) extracted from TeDDI and the scored NLP data sets, we calculate morphological diversity according to both TI and  $\\mathrm{J}_{mm}$ .",
+        "bbox": [
+            112,
+            429,
+            485,
+            492
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Morphology Typological index  $\\left(\\mathbf{TI}_{\\mathit{morph}}\\right)$  We adapt the measure proposed by Ponti et al. (2020) to the text-based features (mean word length). Each bin of the mean word length values is a feature and the number of languages that fall in a given bin are the counts of feature values. In other words, the mean word length becomes a vector of binary values, 1 for the languages that are in the bin and 0 for all the other languages in the sample. The rest of the calculation is the same as in  $\\mathrm{TI}_{\\mathrm{syn}}$ .",
+        "bbox": [
+            112,
+            495,
+            485,
+            656
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Morphology Jaccard (J_mm_morph) Similarly to  $J_{-}mm_{-}\\mathrm{syn}$ , we calculate the Jaccard score by comparing the distributions of the mean word length: TeDDi vs. a given NLP data set.",
+        "bbox": [
+            112,
+            659,
+            485,
+            722
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "6 Findings",
+        "text_level": 1,
+        "bbox": [
+            112,
+            733,
+            225,
+            750
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 1 lists all the reviewed data sets with all the measures of linguistic diversity. The colour scale of the cells represents the relative ranking of data sets according to each measure separately. TeDDI data set obtains the highest diversity scores at both levels (syntax and morphology) using the TI measure. This confirms the role of these resources as the current reference regarding linguistic diversity.",
+        "bbox": [
+            112,
+            757,
+            485,
+            885
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "TI and  $\\mathbf{J}_{mm}$  are consistent The rankings of data sets according to the  $\\mathbf{J}_{mm}$  score are very similar",
+        "bbox": [
+            112,
+            889,
+            485,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "to those obtained with the TI score when the syntactic features are used. The agreement between the two measures is somewhat lower in the case of morphological features, but still rather high. The consistency between the two measures is not a trivial outcome given the entirely different approaches behind them. We can thus take this agreement as a validation of both measures. The main advantage of  $\\mathrm{J}_{mm}$  compared to TI is its transparency regarding the kinds of languages that are missing. The difference with respect to the reference is visible at the level of features indicating the values that need to be added or removed to improve the diversity.",
+        "bbox": [
+            505,
+            354,
+            884,
+            564
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Diversity rankings of NLP data sets The highest rankings appear split between the two structural levels. Bible 100 (Christodouloupoulos and Steedman, 2015) and XTREM (Hu et al., 2020) are the two most syntactically diverse data sets, while their morphological diversity is moderate to low. The Bible data set contains mostly non Indo-European languages, while the collection criteria for the XTREM data set was to maximise diversity. On the other hand, Universal Dependencies (UD, Nivre et al. (2020), which are often seen as especially biased towards European languages, show the best morphological, but a moderate syntactic diversity. XCOPA (Ponti et al., 2020) and TyDiQA (Clark et al., 2020) are data sets containing relatively few languages, but designed to maximise linguistic diversity. They are both highly ranked on  $3/4$  measures (two syntactic and one morphological). Contrary to this, the linguistic diversity ranking of one of the most popular benchmarks that contain manual labels for several downstream tasks, XGLUE (Liang et al., 2020; Wang et al., 2019) is",
+        "bbox": [
+            507,
+            567,
+            884,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "3373",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/aa791db69ee4a5869781157fef0028d7c50b61b5542c400dcfa42430f1229f87.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            84,
+            361,
+            202
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/b3d760642ba5bc28c41abcae5b726d3761766985d672fc626e045a99b8b167d6.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            376,
+            83,
+            620,
+            200
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/7ce5b4f51df11ee765152472d271340ed9e46810fd2fae650336e8088db430ce.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            638,
+            85,
+            880,
+            202
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/6b0343e5da51fe91d1f850d541582fbbbeceffdc75c430e5c922b609007b05b0.jpg",
+        "image_caption": [
+            "Figure 3: Union and intersection between the distributions of the mean word length in TeDDi and NLP data sets."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            115,
+            215,
+            364,
+            334
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/a635d77ac10849d7ce14fc95cbbe6f453e4efae11ec412413b519f6c05fe8883.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            374,
+            215,
+            621,
+            334
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/192ccbdb7a0fcb090534cd9875c0e6f57289b937c3ed5f4b2222ff62bdbd875b.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            633,
+            215,
+            880,
+            335
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "consistently low. XQuAD (Artetxe et al., 2020; Rajpurkar et al., 2016) fairs a little better, but it is still one of the least diverse data sets. The XNLI data set (Conneau et al., 2018; Bowman et al., 2015; Williams et al., 2018), which is compiled with the goal of spanning language families and which includes some low resource languages, remains of moderate linguistic diversity according to all measures. It is curious to see that the number of languages or even languages families included in a data set does not ensure a high linguistic diversity. For example, the mBERT $^{14}$  data set contains 97 languages in 15 language families, but it turns out to be less diverse than smaller data sets such as XCOPA (on  $\\mathrm{TI}_{\\mathrm{syn}}$ ,  $\\mathrm{J}_{\\mathrm{mm\\_syn}}$  and  $\\mathrm{J}_{\\mathrm{mm\\_morph}}$ ) and TyDiQA (on  $\\mathrm{TI}_{\\mathrm{syn}}$ ,  $\\mathrm{J}_{\\mathrm{mm\\_syn}}$  and  $\\mathrm{TI}_{\\mathrm{morph}}$ ). The strategy of including the top 100 languages according to the size of their Wikipedia content (plus Thai and Mongolian), does not result in high diversity.",
+        "bbox": [
+            112,
+            385,
+            489,
+            692
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Underrepresented language types Figure 3 is a visualisation of the  $J_{mm\\_ morph}$  score<sup>15</sup> for some of the data sets showing the overlap and differences with the reference (TeDDi). The recurrent difference is whether a data set includes languages with long words or not (mean length  $>8$ ). Those that contain at least some languages with long words (UD, XCOPA) score much better on  $J_{mm\\_ morph}$  than those that remain completely on",
+        "bbox": [
+            112,
+            697,
+            489,
+            845
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "the short-middle side (EXTREME, XGLUE, TyDiQA, mBERT). The second important factor that leads to lower scores is a strong peak of the distribution indicating a bias towards one of the length bins (EXTREME, XGLUE, mBERT). The third factor is a different (\"wrong\") shape of the distribution (TyDiQA). The data set that diverges the most is EXTREME, exhibiting all three factors of disagreement.",
+        "bbox": [
+            507,
+            385,
+            884,
+            529
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "The information about what kinds of languages are missing in a data set can be used to adjust language sampling and improve diversity. This is relatively straightforward when we deal with a single feature such as the mean word length. For example, the diversity of the mBERT language sample would be improved if the number of languages with a mean word length between 3 and 4 is reduced (by removing a given number of randomly selected languages). Instead of these languages, one should add a given number of languages with a mean word length greater than 7. It is not obvious where to look and how to find such languages (beyond the TeDDI sample), but knowing that they are needed might motivate such searches. Multifeature scores (such as feature entropy) could specify the needed languages more precisely, but they would require an optimisation method to ensure that a newly added language increases indeed the diversity score. It might happen, for instance, that we want to increase the count on one feature value but not on another. In this case, we need a language that has 1 on the desired feature value but 0 on the features that we do not want to change. Devising",
+        "bbox": [
+            507,
+            535,
+            884,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{14}$ https://github.com/google-research/bert/blob/master/multilingual.md",
+        "bbox": [
+            112,
+            857,
+            487,
+            882
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>15</sup>We show the morphological diversity for convenience since visualising 103 syntactic features would require additional adaptations.",
+        "bbox": [
+            112,
+            882,
+            487,
+            920
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "3374",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "such a method is beyond the scope of the current paper, but it is a clear next step for future work.",
+        "bbox": [
+            112,
+            84,
+            485,
+            116
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Overall, it seems that the right-hand side of the mean word length scale remains rather scarcely represented in all data sets, including the TeDDi sample itself. In future data collection, more effort should be put into representing languages with long words, especially because most of them are endangered. There are 12 languages in the TeDDI sample with a mean word length of over 7. If we localise them in Figure 1, we can see that ten of them are classified as extinct, endangered or vulnerable: Apurina (apu), Chukchi (ckt), Kalaallisut (kal), Kayardild (gyd), Makah (myh), Martuthunira (vma), Plains Cree (crk), Ngiyambaa (wyb), Wichita (wic) and Yagua (yad). Only two of these languages, Luvale (lue) and Zulu (zul) are safe.",
+        "bbox": [
+            115,
+            118,
+            489,
+            359
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "7 Discussion",
+        "text_level": 1,
+        "bbox": [
+            112,
+            375,
+            240,
+            390
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Our linguistic diversity scores include two kinds of language features (expert features extracted from data bases and the mean word length as a text feature) describing two structural levels (syntax and morphology). Readers not familiar with the details of how expert features are used in NLP might be left wondering whether the use of the mean word length is necessary and whether this measure is a good approximation of morphological types.",
+        "bbox": [
+            112,
+            403,
+            487,
+            546
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Describing the use of expert features in NLP in Section 4, we note that the library lang2vec does not contain any morphological features, although these features are present in linguistic data bases. It is not clear why this is the case, but this means that morphological features are currently not used in NLP to assess linguistic diversity and the distances between languages. One possible reason for omitting morphological features could be the problem of sparsity, which would become even worse with these features leading to even more imputed values. For instance, if we want to study the distribution of 27 morphological features, only 34 languages will have a value for all these features. The values for the thousands of other languages would need to be imputed. This is the main reason why we propose to complement the existing sources of expert features with the mean word length as a value that can be easily calculated for any language on a small sample of text (500 tokens).",
+        "bbox": [
+            115,
+            550,
+            487,
+            870
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "To justify this proposal, we show that an independent measure of morphological complexity  $(C_{WALS})$  and the mean word length are strongly",
+        "bbox": [
+            112,
+            873,
+            487,
+            921
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "correlated, but the score of 0.69 means that the agreement is not perfect. A closer look into these two variables (Table 3 in Appendix B) points to the limitations of both measures, especially concerning the high values. For example, Turkish is the most complex language according to  $C_{WALS}$ , but its mean word length is well under 7. Although the correlation score is high and not due to chance, such aggregate measures remain approximations of the structural properties of languages. Nevertheless, these approximations are useful for tracking and improving linguistic diversity in data sets at the level of precision that is currently possible. Better approximations are certainly achievable in future work. Since our methods are general and can be applied to any set of features, any future improvements in representing linguistic structures can be easily integrated.",
+        "bbox": [
+            507,
+            84,
+            885,
+            374
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "8 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            385,
+            640,
+            400
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We have shown that the linguistic diversity of NLP data sets can be consistently assessed by two independent measures, TI (proposed in previous work) and  $\\mathrm{J}_{mm}$  (proposed in this paper). Both of these measures show that a high number of languages and language families included in a data set is not sufficient to ensure linguistic diversity.",
+        "bbox": [
+            507,
+            410,
+            882,
+            523
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "To make the assessment of linguistic diversity automatic and rather simple, we show that text-based features such as the mean word length can be used as linguistic descriptors. These features can be easily calculated on very small text samples (of length of 500 tokens), overcoming the obstacles posed by the need to extract linguistic features from typological databases.",
+        "bbox": [
+            507,
+            523,
+            882,
+            651
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "An advantage of the  $\\mathrm{J}_{mm}$  score over TI and other previous indicators of linguistic diversity is its capacity to show what kinds of languages are missing in a given data set in comparison to a reference. Assessing popular NLP data sets with this measure revealed that the most underrepresented languages are those with rich morphology. This kind of direct and transparent comparison can improve multilingual NLP coverage in the long run.",
+        "bbox": [
+            507,
+            652,
+            884,
+            797
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            509,
+            809,
+            680,
+            825
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This research is partially supported by the Swiss National Science Foundation (SNSF) grants 176305 and PCEFP1_186841. We thank the anonymous reviewers for their suggestions, which have improved the clarity of the paper.",
+        "bbox": [
+            507,
+            834,
+            882,
+            915
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "3375",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            220,
+            99
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "A limitation of our study is that the two levels of linguistic structures are represented with different features: syntax with expert features from linguistic data bases and morphology with mean word length as a text feature. Our results suggest that the two measures agree more at the level of syntax than at the level of morphology. To draw sound conclusions about the impact of the structural level on the agreement between the two measures, we would need both kinds of features for both levels. While we indirectly compare text and expert features at the level of morphology (via the correlation test), we do not propose syntactic features that could be extracted from text. We focused here on the current gap in the available linguistic features (the lack of morphological features in lang2vec), but devising and validating text-based syntactic features would deserve more attention in future work.",
+        "bbox": [
+            112,
+            109,
+            492,
+            399
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            425,
+            213,
+            439
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Rami Al-Rfou. 2015. *Polyglot: A massive multilingual natural language processing pipeline*. Ph.D. thesis, State University of New York at Stony Brook.",
+            "Mikel Artetxe, Sebastian Ruder, and Dani Yogatama. 2020. On the cross-lingual transferability of monolingual representations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4623–4637, Online. Association for Computational Linguistics.",
+            "Emily M. Bender and Batya Friedman. 2018. Data statements for natural language processing: Toward mitigating system bias and enabling better science. Transactions of the Association for Computational Linguistics, 6:587-604.",
+            "Christian Bentz, Tatyana Ruzsics, Alexander Koplenig, and Tanja Samardzic. 2016. A comparison between morphological complexity measures: typological data vs. language corpora. In Proceedings of the workshop on computational linguistics for linguistic complexity (cl4lc), pages 142-153.",
+            "Aleksandrs Berdicevskis, Căgni Čöltekin, Katharina Ehret, Kilu von Prince, Daniel Ross, Bill Thompson, Chunxiao Yan, Vera Demberg, Gary Lupyan, Taraka Rama, et al. 2018. Using universal dependencies in cross-linguistic complexity research. In Proceedings of the Second Workshop on Universal Dependencies (UDW 2018), pages 8-17.",
+            "Balthasar Bickel, Johanna Nichols, Taras Zakharko, Alena Witzlack-Makarevich, Kristine Hildebrandt, Michael Rießler, Lennart Bierkandt, Fernando Zúñiga, and John B Lowe. 2017. The autotyp typological databases. version 0.1.2."
+        ],
+        "bbox": [
+            115,
+            448,
+            489,
+            920
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. 2015. A large annotated corpus for learning natural language inference. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 632-642, Lisbon, Portugal. Association for Computational Linguistics.",
+            "Christos Christodouloupoulos and Mark Steedman. 2015. A massively parallel corpus: the bible in 100 languages. Language Resources and Evaluation, 49(2):375-395.",
+            "Jonathan H. Clark, Eunsol Choi, Michael Collins, Dan Garrette, Tom Kwiatkowski, Vitaly Nikolaev, and Jennimaria Palomaki. 2020. TyDi QA: A benchmark for information-seeking question answering in typologically diverse languages. Transactions of the Association for Computational Linguistics, 8:454-470.",
+            "Bernard Comrie, Matthew S. Dryer, David Gil, and Martin Haspelmath. 2013. Introduction. In Matthew S. Dryer and Martin Haspelmath, editors, The World Atlas of Language Structures Online. Max Planck Institute for Evolutionary Anthropology, Leipzig.",
+            "Alexis Conneau, Rudy Rinott, Guillaume Lample, Adina Williams, Samuel Bowman, Holger Schwenk, and Veselin Stoyanov. 2018. XNLI: Evaluating cross-lingual sentence representations. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2475-2485, Brussels, Belgium. Association for Computational Linguistics.",
+            "Wietse de Vries, Martijn Wieling, and Malvina Nissim. 2022. Make the best of cross-lingual transfer: Evidence from POS tagging with over 100 languages. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7676-7685, Dublin, Ireland. Association for Computational Linguistics.",
+            "Matthew S. Dryer and Martin Haspelmath, editors. 2013. WALS Online. Max Planck Institute for Evolutionary Anthropology, Leipzig.",
+            "Gregory Grefenstette. 1994. Explorations in Automatic Thesaurus Discovery. Kluwer Academic Publishers, USA.",
+            "Harald Hammarström, Robert Forkel, and Martin Haspelmath. 2018. Glottolog 3.3. Leipzig.",
+            "Martin Haspelmath. 2007. Pre-established categories don't exist: Consequences for language description and typology. Linguistic Typology, 11(1):119-132.",
+            "Junjie Hu, Sebastian Ruder, Aditya Siddhant, Graham Neubig, Orhan First, and Melvin Johnson. 2020. XTREME: A massively multilingual multitask benchmark for evaluating cross-lingual generalisation. In Proceedings of the 37th International"
+        ],
+        "bbox": [
+            510,
+            85,
+            885,
+            920
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "3376",
+        "bbox": [
+            480,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, volume 119 of Proceedings of Machine Learning Research, pages 4411-4421. PMLR.",
+            "P. Jaccard. 1912. The distribution of the flora in the alpine zone.1. New Phytologist, 11:37-50.",
+            "Pratik Joshi, Sebastin Santy, Amar Budhiraja, Kalika Bali, and Monojit Choudhury. 2020. The state and fate of linguistic diversity and inclusion in the NLP world. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6282-6293, Online. Association for Computational Linguistics.",
+            "Kimmo Kettunen. 2014. Can type-token ratio be used to show morphological complexity of languages? Journal of Quantitative Linguistics, 21(3):223-245.",
+            "Julia Kreutzer, Isaac Caswell, Lisa Wang, Ahsan Wahab, Daan van Esch, Nasanbayar Ulzii-Orshikh, Alahsera Tapo, Nishant Subramani, Artem Sokolov, Claytone Sikasote, Monang Setyawan, Supheakmungkol Sarin, Sokhar Samb, Benoit Sagot, Clara Rivera, Annette Rios, Isabel Papadimitriou, Salomey Osei, Pedro Ortiz Suárez, Iroro Orife, Kelechi Ogueji, Andre Niyongabo Rubungo, Toan Q. Nguyen, Mathias Müller, André Müller, Shamsuddeen Hassan Muhammad, Nanda Muhammad, Ayanda Mnyakeni, Jamshidbek Mirzakhalov, Tapiwanashe Matangira, Colin Leong, Nze Lawson, Sneha Kudugunta, Yacine Jernite, Mathias Jenny, Orhan First, Bonaventure F. P. Dossou, Sakhile Dlamini, Nisansa de Silva, Sakine Cabuk Balli, Stella Biderman, Alessia Battisti, Ahmed Baruwa, Ankur Bapna, Pallavi Baljekar, Israel Abebe Azime, Ayodele Awokoya, Duygu Ataman, Orevaoghene Ahia, Oghenefego Ahia, Sweta Agrawal, and Mofetoluwa Adeyemi. 2021. Quality at a glance: An audit of web-crawled multilingual datasets.",
+            "Anne Lauscher, Vinit Ravishankar, Ivan Vulic, and Goran Glavaš. 2020. From zero to hero: On the limitations of zero-shot language transfer with multilingual Transformers. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4483-4499, Online. Association for Computational Linguistics.",
+            "Quentin Lhoest, Albert Villanova del Moral, Yacine Jernite, Abhishek Thakur, Patrick von Platen, Suraj Patil, Julien Chaumont, Mariama Drame, Julien Plu, Lewis Tunstall, Joe Davison, Mario Šaško, Gunjan Chhablani, Bhavitvya Malik, Simon Brandeis, Teven Le Scao, Victor Sanh, Canwen Xu, Nicolas Patry, Angelina McMillan-Major, Philipp Schmid, Sylvain Gugger, Clément Delangue, Theo Matussière, Lysandre Debut, Stas Bekman, Pierrick Cistac, Thibault Goehringer, Victor Mustar, François Lagunas, Alexander Rush, and Thomas Wolf. 2021. Datasets: A community library for natural language processing. In Proceedings of the 2021 Conference"
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            919
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "on Empirical Methods in Natural Language Processing: System Demonstrations, pages 175-184, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Yaobo Liang, Nan Duan, Yeyun Gong, Ning Wu, Fenfei Guo, Weizhen Qi, Ming Gong, Linjun Shou, Daxin Jiang, Guihong Cao, Xiaodong Fan, Ruofei Zhang, Rahul Agrawal, Edward Cui, Sining Wei, Taroon Bharti, Ying Qiao, Jiun-Hung Chen, Winnie Wu, Shuguang Liu, Fan Yang, Daniel Campos, Rangan Majumder, and Ming Zhou. 2020. XGLUE: A new benchmark dataset for cross-lingual pretraining, understanding and generation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6008-6018, Online. Association for Computational Linguistics.",
+            "Yu-Hsiang Lin, Chian-Yu Chen, Jean Lee, Zirui Li, Yuyan Zhang, Mengzhou Xia, Shruti Rijhwani, Junxian He, Zhisong Zhang, Xuezhe Ma, Antonios Anastasopoulos, Patrick Littell, and Graham Neubig. 2019. Choosing transfer languages for cross-lingual learning. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3125-3135, Florence, Italy. Association for Computational Linguistics.",
+            "Pierre Lison and Jörg Tiedemann. 2016. Opensubtitles2016: Extracting large parallel corpora from movie and tv subtitles. In Proceedings from LREC 2016, pages 923-929. European Language Resources Association.",
+            "Patrick Littell, David R. Mortensen, Ke Lin, Katherine Kairis, Carlisle Turner, and Lori Levin. 2017. URIEL and lang2vec: Representing languages as typological, geographical, and phylogenetic vectors. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, pages 8-14, Valencia, Spain. Association for Computational Linguistics.",
+            "Thomas Mayer and Michael Cysouw. 2014. Creating a massively parallel bible corpus. In Proceedings of the International Conference on Language Resources and Evaluation (LREC), pages 3158-3163.",
+            "Steven Moran. 2016. The ACQDIV database: Min(d)ing the ambient language. In Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC'16), pages 4423-4429, Portoorž, Slovenia. European Language Resources Association (ELRA).",
+            "Steven Moran, Christian Bentz, Ximena Gutierrez-Vasques, Olga Pelloni, and Tanja Samardzic. 2022. TeDDi sample: Text data diversity sample for language comparison and multilingual NLP. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 1150-1158, Marseille, France. European Language Resources Association."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            919
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "3377",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Steven Moran and Michael Cysouw. 2018. The Unicode cookbook for linguists. Number 10 in Translation and Multilingual Natural Language Processing. Language Science Press, Berlin.",
+            "Joakim Nivre, Marie-Catherine de Marneffe, Filip Ginter, Jan Hajic, Christopher D. Manning, Sampo Pyysalo, Sebastian Schuster, Francis Tyers, and Daniel Zeman. 2020. Universal Dependencies v2: An evergrowing multilingual treebank collection. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 4034-4043, Marseille, France. European Language Resources Association.",
+            "Olga Pelloni, Anastassia Shaitarova, and Tanja Samardzic. 2022. Subword evenness (SuE) as a predictor of cross-lingual transfer to low-resource languages. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 7428-7445, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.",
+            "Telmo Pires, Eva Schlinger, and Dan Garrette. 2019. How multilingual is multilingual BERT? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4996-5001, Florence, Italy. Association for Computational Linguistics.",
+            "Edoardo Maria Ponti, Goran Glavaš, Olga Majewska, Qianchu Liu, Ivan Vulić, and Anna Korhonen. 2020. XCOPA: A multilingual dataset for causal commonsense reasoning. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2362-2376, Online. Association for Computational Linguistics.",
+            "Edoardo Maria Ponti, Helen O'Horan, Yevgeni Berzak, Ivan Vulic, Roi Reichart, Thierry Poibeau, Ekaterina Shutova, and Anna Korhonen. 2019. Modeling language variation and universals: A survey on typological linguistics for natural language processing. Computational Linguistics, 45(3):559-601.",
+            "Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ questions for machine comprehension of text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383-2392, Austin, Texas. Association for Computational Linguistics.",
+            "Tanja Samardžić, Ximena Gutierrez-Vasques, Rob van der Goot, Max Müller-Eberstein, Olga Pelloni, and Barbara Plank. 2022. On language spaces, scales and cross-lingual transfer of UD parsers. In Proceedings of the 26th Conference on Computational Natural Language Learning (CoNLL), pages 266–281, Abu Dhabi, United Arab Emirates (Hybrid). Association for Computational Linguistics.",
+            "Richard Sproat and Alexander Gutkin. 2021. The Taxonomy of Writing Systems: How to Measure How Logographic a System Is. Computational Linguistics, 47(3):477-528."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            920
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Sabine Stoll and Balthasar Bickel. 2013. Capturing diversity in language acquisition research. In Balthasar Bickel, Lenore A. Grenoble, David A. Peterson, and Alan Timberlake, editors, Language typology and historical contingency: studies in honor of Johanna Nichols, pages 195-260. Benjamins, Amsterdam. [pre-print available at http://www.psycholinguistics.uzh.ch/stoll/publications/stollbickel.sampling2012rev.pdf].",
+            "T. T Tanimoto. 1958. An elementary mathematical theory of classification and prediction. International Business Machines Corporation.",
+            "Reut Tsarfaty, Djamé Seddah, Sandra Kübler, and Joakim Nivre. 2013. Parsing morphologically rich languages: Introduction to the special issue. Computational Linguistics, 39(1):15-22.",
+            "Iulia Turc, Kenton Lee, Jacob Eisenstein, Ming-Wei Chang, and Kristina Toutanova. 2021. Revisiting the primacy of english in zero-shot cross-lingual transfer.",
+            "Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. 2019. GLUE: A multi-task benchmark and analysis platform for natural language understanding. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net.",
+            "Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. A broad-coverage challenge corpus for sentence understanding through inference. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1112-1122, New Orleans, Louisiana. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            621
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "3378",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A Mean Word Length Correlation between Different Sample Size",
+        "text_level": 1,
+        "bbox": [
+            114,
+            83,
+            695,
+            99
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "image",
+        "img_path": "images/154462964f9a90dc5a5d1ee6832b4508d0a7d3437df4a8d0954cb1ba47c0d136.jpg",
+        "image_caption": [
+            "Figure 4: Mean word length measures at different text sizes in TeDDi. The languages on the x-axis are sorted according to the increasing value calculated on the biggest sample (10K). The values in the two smaller samples (2K and 500) depart very little from the main trend."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            123,
+            880,
+            350
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "To make sure that the stability across different sample sizes suggested by Figure 4 is not a mere consequence of a relatively small range of variation, we perform correlation tests between different samples and in comparison to other measures (TTR and unigram entropy (H)). Table 2 shows that the ranks of languages change considerably less across different sample sizes when considering the mean word length than in the other two measures.",
+        "bbox": [
+            112,
+            417,
+            882,
+            498
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "table",
+        "img_path": "images/c7ff99446e99ca6d27e88efe576d8f9ad0db57ae15d51685799211e49e3b8e63.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Samples</td><td>MWL</td><td>H</td><td>TTR</td></tr><tr><td>500 tokens vs. max.</td><td>0.99</td><td>0.85</td><td>0.84</td></tr><tr><td>2K tokens vs. max</td><td>0.99</td><td>0.95</td><td>0.94</td></tr></table>",
+        "bbox": [
+            322,
+            508,
+            673,
+            564
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Spearman rank correlation showing how much rankings of languages change with text measures taken on random samples of different size.",
+        "bbox": [
+            112,
+            574,
+            882,
+            602
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "3379",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "table",
+        "img_path": "images/650fca991571e20dbe29147f76fc8ee2b6d7b99b366d6961cbce9ab86adc1bd9.jpg",
+        "table_caption": [
+            "B Word length and morphological complexity"
+        ],
+        "table_footnote": [
+            "Spearmann correlation  $\\rho = 0.69$"
+        ],
+        "table_body": "<table><tr><td>ISO396-3</td><td>Name</td><td>MWL</td><td>CwALS</td></tr><tr><td>abk</td><td>Abkhazian</td><td>7.17</td><td>0.62</td></tr><tr><td>apu</td><td>Apurinã</td><td>7.67</td><td>0.60</td></tr><tr><td>arz</td><td>Egyptian Arabic</td><td>4.44</td><td>0.49</td></tr><tr><td>bsn</td><td>Barasana-Eduria</td><td>6.02</td><td>0.69</td></tr><tr><td>ckt</td><td>Chukchi</td><td>8.45</td><td>0.50</td></tr><tr><td>deu</td><td>German</td><td>4.87</td><td>0.55</td></tr><tr><td>ell</td><td>Modern Greek</td><td>4.72</td><td>0.53</td></tr><tr><td>eng</td><td>English</td><td>4.18</td><td>0.42</td></tr><tr><td>eus</td><td>Basque</td><td>5.70</td><td>0.64</td></tr><tr><td>fin</td><td>Finnish</td><td>6.23</td><td>0.66</td></tr><tr><td>fra</td><td>French</td><td>4.41</td><td>0.45</td></tr><tr><td>hae</td><td>Eastern Oromo</td><td>5.91</td><td>0.53</td></tr><tr><td>hau</td><td>Hausa</td><td>4.08</td><td>0.38</td></tr><tr><td>heb</td><td>Modern Hebrew</td><td>3.94</td><td>0.54</td></tr><tr><td>ind</td><td>Indonesian</td><td>5.42</td><td>0.40</td></tr><tr><td>kan</td><td>Kannada</td><td>5.22</td><td>0.65</td></tr><tr><td>kat</td><td>Georgian</td><td>4.78</td><td>0.50</td></tr><tr><td>khk</td><td>Halh Mongolian</td><td>5.66</td><td>0.53</td></tr><tr><td>kut</td><td>Kutenai</td><td>4.60</td><td>0.37</td></tr><tr><td>lvk</td><td>Lavukaleve</td><td>4.77</td><td>0.67</td></tr><tr><td>qvi</td><td>Imbabura Highland Quichua</td><td>8.18</td><td>0.71</td></tr><tr><td>rus</td><td>Russian</td><td>4.79</td><td>0.52</td></tr><tr><td>spa</td><td>Spanish</td><td>4.37</td><td>0.45</td></tr><tr><td>swh</td><td>Swahili</td><td>5.72</td><td>0.71</td></tr><tr><td>tur</td><td>Turkish</td><td>6.07</td><td>0.76</td></tr><tr><td>vie</td><td>Vietnamese</td><td>3.20</td><td>0.21</td></tr><tr><td>yaq</td><td>Yaqui</td><td>5.31</td><td>0.57</td></tr><tr><td>yor</td><td>Yoruba</td><td>3.52</td><td>0.25</td></tr></table>",
+        "bbox": [
+            114,
+            118,
+            680,
+            590
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Mean Word length (MWL) and morphological complexity measure ( $C_{WALS}$ ) in the subset of TeDDi languages for which 26 WALS morphology features are known.",
+        "bbox": [
+            112,
+            615,
+            882,
+            646
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "page_number",
+        "text": "3380",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "table",
+        "img_path": "images/b0646013ea81a24de545ebb28d1e38fef37a47a5ab69b1ba23e18c20a1012eb8.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Chapter</td><td>Name</td><td>Categories</td><td>Transformation</td><td>Final Values</td></tr><tr><td>22A</td><td>Inflectional Synthesis</td><td>7 (ordinal)</td><td>none</td><td>1-7</td></tr><tr><td>26A</td><td>Prefixing vs. Suffixing in Inflectional Morphology</td><td>6 (non-ordinal)</td><td>binarization</td><td>0-1</td></tr><tr><td>27A</td><td>Reduplication</td><td>3 (non-ordinal)</td><td>binarization</td><td>0-1</td></tr><tr><td>28A</td><td>Case Syncretism</td><td>4 (ordinal)</td><td>reorder</td><td>1-4</td></tr><tr><td>29A</td><td>Syncretism in Verbal Per-son/Number marking</td><td>3 (ordinal)</td><td>none</td><td>1-3</td></tr><tr><td>30A</td><td>Number of Genders</td><td>5 (ordinal)</td><td>none</td><td>1-5</td></tr><tr><td>33A</td><td>Coding of Nominal Plurality</td><td>9 (partially ordinal)</td><td>binarization</td><td>0-1</td></tr><tr><td>34A</td><td>Occurrence of Nominal Plurality</td><td>6 (ordinal)</td><td>none</td><td>1-6</td></tr><tr><td>49A</td><td>Number of Cases</td><td>9 (ordinal)</td><td>remove</td><td>1-8</td></tr><tr><td>51A</td><td>Position of Case Affixes</td><td>9 (non-ordinal)</td><td>binarization</td><td>0-1</td></tr><tr><td>57A</td><td>Position of Pronominal Posses-sive Affixes</td><td>4 (non-ordinal)</td><td>binarization</td><td>0-1</td></tr><tr><td>59A</td><td>Possessive Classification</td><td>4 (ordinal)</td><td>none</td><td>1-4</td></tr><tr><td>65A</td><td>Perfective/Imperfective Aspect</td><td>binary</td><td>none</td><td>0-1</td></tr><tr><td>66A</td><td>The Past Tense</td><td>4 (ordinal)</td><td>reorder</td><td>1-4</td></tr><tr><td>67A</td><td>The Future Tense</td><td>binary</td><td>none</td><td>0-1</td></tr><tr><td>69A</td><td>Position of Tense/Aspect Affixes</td><td>5 (non-ordinal)</td><td>binarization</td><td>0-1</td></tr><tr><td>70A</td><td>The Morphological Imperative</td><td>5 (partially ordinal)</td><td>recategorization</td><td>1-4</td></tr><tr><td>73A</td><td>The Optative</td><td>binary</td><td>none</td><td>0-1</td></tr><tr><td>74A</td><td>Situational Possibility</td><td>3 (non-ordinal)</td><td>binarization</td><td>0-1</td></tr><tr><td>75A</td><td>Epistemic Possibility</td><td>3 (non-ordinal)</td><td>binarization</td><td>0-1</td></tr><tr><td>78A</td><td>Coding of Evidentiality</td><td>6 (non-ordinal)</td><td>binarization</td><td>0-1</td></tr><tr><td>94A</td><td>Subordination</td><td>5 (non-ordinal)</td><td>binarization</td><td>0-1</td></tr><tr><td>101A</td><td>Expression of Pronominal Sub-jects</td><td>6 (non-ordinal)</td><td>binarization</td><td>0-1</td></tr><tr><td>102A</td><td>Verbal Person Marking</td><td>5 (partially ordinal)</td><td>recategorization</td><td>1-3</td></tr><tr><td>111A</td><td>Nonperiphrastic Causative Constructions</td><td>4 (non-ordinal)</td><td>binarization</td><td>0-1</td></tr><tr><td>112A</td><td>Negative Morphemes</td><td>6 (non-ordinal)</td><td>binarization</td><td>0-1</td></tr></table>",
+        "bbox": [
+            114,
+            211,
+            885,
+            734
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "Table 4: Subset of WALS features that we use for characterizing the morphological complexity of languages. The column \"Final Values\" gives the range of values each feature can take after transformations were performed to the original values (Bentz et al., 2016)",
+        "bbox": [
+            112,
+            741,
+            884,
+            785
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "page_number",
+        "text": "3381",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "C Word Length Adjustments for Logographic Scripts",
+        "text_level": 1,
+        "bbox": [
+            114,
+            83,
+            593,
+            101
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "table",
+        "img_path": "images/39d34dd19e2995e8329dd7e6a16ae784fadb1fa24169d00f741ca1a1e3fff6a1.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Name and main references</td><td>N(L)</td><td>N(F)</td><td>TI_syn</td><td>Jmm_syn</td><td>TI_morph</td><td>Jmm_morph</td></tr><tr><td>Universal Dependencies (UD)</td><td>106*</td><td>20*</td><td>0.567</td><td>0.736</td><td>0.337</td><td>0.665</td></tr><tr><td>Bible 100</td><td>103*</td><td>30*</td><td>0.649</td><td>0.811</td><td>0.302</td><td>0.617</td></tr><tr><td>mBERT</td><td>97*</td><td>15*</td><td>0.559</td><td>0.710</td><td>0.316</td><td>0.617</td></tr><tr><td>XTREME</td><td>40</td><td>14</td><td>0.612</td><td>0.775</td><td>0.311</td><td>0.471</td></tr><tr><td>XGLUE</td><td>19</td><td>7*</td><td>0.517</td><td>0.674</td><td>0.297</td><td>0.580</td></tr><tr><td>XNLI</td><td>15</td><td>7*</td><td>0.557</td><td>0.711</td><td>0.321</td><td>0.704</td></tr><tr><td>XCOPA</td><td>11</td><td>11</td><td>0.586</td><td>0.737</td><td>0.336</td><td>0.634</td></tr><tr><td>TyDiQA</td><td>11</td><td>10</td><td>0.626</td><td>0.751</td><td>0.343</td><td>0.552</td></tr><tr><td>XQuAD</td><td>12*</td><td>6*</td><td>0.523</td><td>0.680</td><td>0.318</td><td>0.634</td></tr><tr><td>TeDDi</td><td>89</td><td>51</td><td>0.706</td><td>-</td><td>0.361</td><td>-</td></tr></table>",
+        "bbox": [
+            121,
+            118,
+            875,
+            300
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "text",
+        "text": "Table 5: Diversity of multilingual NLP data sets with adjustments for logographic scripts. Compared to the main results in Table 1, all  $\\mathrm{TI}_{\\mathrm{morph}}$  scores are slightly decreased and  $\\mathbf{J}_{mm\\_ morph}$  slightly increased. The rankings of the t are mostly preserved, with the exception of XNLI, whose  $\\mathbf{J}_{mm\\_ morph}$  ranking improves.",
+        "bbox": [
+            112,
+            310,
+            882,
+            355
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "page_number",
+        "text": "3382",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 15
+    }
+]

2024/A Measure for Transparent Comparison of Linguistic Diversity in Multilingual NLP Data Sets/7dee6ef6-08dd-4565-9a55-2f0240bdda97_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e01d86bc58a7429ff3f99ae922b420d338ff094bb7e3c70f4b3fa621f9e950f3
+size 1066219

2024/A Measure for Transparent Comparison of Linguistic Diversity in Multilingual NLP Data Sets/full.md

@@ -0,0 +1,325 @@
+# A Measure for Transparent Comparison of Linguistic Diversity in Multilingual NLP Data Sets
+
+Tanja Samardžić
+
+Language and Space Lab
+
+University of Zurich
+
+tanja.samardzic@uzh.ch
+
+Ximena Gutierrez
+
+CEIICH
+
+Universidad Nacional
+
+Autónoma de México
+
+Christian Bentz
+
+Dept. of General Linguistics
+
+Eberhard-Karls-Universität
+
+Tübingen
+
+Steven Moran
+
+Laboratory of Language Evolution
+University of Neuchâtel
+
+Olga Pelloni
+
+Telepathy Labs GmbH
+
+Zurich, Switzerland
+
+# Abstract
+
+Typologically diverse benchmarks are increasingly created to track the progress achieved in multilingual NLP. Linguistic diversity of these data sets is typically measured as the number of languages or language families included in the sample, but such measures do not consider structural properties of the included languages. In this paper, we propose assessing linguistic diversity of a data set against a reference language sample as a means of maximising linguistic diversity in the long run. We represent languages as sets of features and apply a version of the Jaccard index  $(J_{mm})$  suitable for comparing sets of measures. In addition to the features extracted from typological data bases, we propose an automatic text-based measure, which can be used as a means of overcoming the well-known problem of data sparsity in manually collected features. Our diversity score is interpretable in terms of linguistic features and can identify the types of languages that are not represented in a data set. Using our method, we analyse a range of popular multilingual data sets (UD, Bible100, mBERT, XTREME, XGLUE, XNLI, XCOPA, TyDiQA, XQuAD). In addition to ranking these data sets, we find, for example, that (poly)synthetic languages are missing in almost all of them.
+
+# 1 Introduction
+
+Data sets for training and testing NLP models are increasingly multilingual and aimed at broad linguistic coverage. These data sets are often claimed to represent a typologically diverse sample, including low-resource and endangered languages.
+
+Linguistic diversity is typically described as the number of languages included in the data set, yet less often as the number of language families to which these languages belong. Both counts indicate a level of linguistic diversity: the more languages and families, the more diversity. But how
+
+do we know that included languages are indeed different? How can we define a desired or optimal diversity to set as a goal when composing multilingual data sets? These questions need to be addressed if our goal is to know how NLP technology generalises across diverse languages, without testing it on each single language (even if we had the necessary data for all languages).
+
+The aim of this paper is to initiate and facilitate comparisons between multilingual NLP data sets with respect to a linguistic diversity reference. For this, we propose a measure of linguistic diversity and a method of comparison that identifies what kinds of linguistic features are missing. As an initial reference, we rely on a predefined sample of languages — the 100-language-sample (100L) selected by the Word Atlas of Language Structures (WALS; Comrie et al. (2013)) to represent geographic and phylogenetic diversity. As a comparison method, we formulate a version of the Jaccard index suitable for comparing measures. This measure allows us to quantify the distance between the observed and the reference diversity in terms of linguistic features, showing not only how diverse language samples are but also what kinds of linguistic phenomena are not represented in a given sample. To facilitate automatic extraction of linguistic features needed for assessing linguistic diversity, we complement the information from linguistic data bases with relevant text statistics.
+
+Our proposals are intended to help researchers make informed choices when designing a multilingual data set. Representing a wider spectrum of linguistic diversity is not only a way to improve the cross-linguistic generalisation of NLP technology, but also a way to deal with biases against low-resource languages, which are harder to represent and thus more likely to be left behind (Joshi et al., 2020).
+
+![](images/eb7d8c18bf9e049512d0759f01f3135b27f9ce22c7df6c433a234891f914192e.jpg)  
+Figure 1: Languages in the WALS 100L sample with their endangerment status.
+
+# 2 Background and Related Work
+
+Evaluating the linguistic diversity of data sets relies on comparable descriptions of languages. For instance, the (approximate) number of speakers is an attribute whose value can be found and compared for all registered languages. This attribute, however, does not describe the structure of languages. An example of a structural attribute would be the presence or the absence of adjectives in a language. To establish the value of this attribute for any language, we need a universal definition of what an adjective is. It turns out that such universal definitions are hard to formulate in a principled way (Haspelmath, 2007), which makes it hard to define objective measures of how similar or dissimilar any two languages are.
+
+The most widely accepted method for comparing languages relies on genealogical classification: given a phylogenetic tree, we consider languages located in the same region of the tree to be similar. This method currently prevails in NLP (cf. the work discussed in Section 6). Typically, we regard languages that belong to the same family to be similar. To know which language belongs to which family, we turn to popular authorities such as WALS (Dryer and Haspelmath, 2013) or Glottolog (Hammarström et al., 2018). However, language families can be too broad for a meaningful comparison as they include typologically very different languages. For instance, English and Armenian belong to the same family, Indo-European, but are
+
+vastly different in terms of their phoneme inventories, morphology, and word order.
+
+Another possibility to compare languages, starting to be used in NLP only recently, is to rely on grammatical features available in the WALS data base. This is a comprehensive source of information about linguistic structures but still rather sparsely populated; feature values are often known for only a few languages. $^{1}$  Together with other typological data bases, WALS is included in URIEL, an aggregated and standardised source of language features for various NLP uses. Ponti et al. (2020) propose a diversity score using the features from URIEL (Littell et al., 2017). The score is called typology index and it is calculated as the entropy of feature values (averaged per data set). $^{2}$  In other NLP work, grammatical features (usually termed typological) are used for other purposes, such as predicting the features (Ponti et al., 2019) rather than using them for language sampling in creating multilingual data sets. Moran (2016) use WALS and AUTOTYP features (Stoll and Bickel, 2013) to compose a sample of 10 maximally diverse languages for a corpus-based study of language acquisition.
+
+Finally, languages can be described using features derived from various text statistics, but such features are not commonly used for language sam
+
+pling. Type-token ratio (TTR) or unigram entropy of a text have been shown to correlate with grammar-based morphological complexity measures (Kettunen, 2014; Bentz et al., 2016). Many other methods have been proposed for assessing linguistic complexity using text statistics (see, for instance, Berdicevskis et al. (2018)). All of these measures can, in principle, be used for describing and comparing languages although such comparisons might seem counter-intuitive and hard to interpret in terms of genealogical classification. On the other hand, these features might complement usual descriptions of languages while being more directly relevant to text processing and NLP.
+
+Transfer learning created a new need for nuanced languages comparison for NLP. While models can now be transferred across languages with zero-shot or few-shot learning (Pires et al., 2019), the success of the transfer might depend on the differences between languages. Lin et al. (2019) propose a range of measures that can be used in order to choose the best transfer language, which they divide into data-dependent (data size, token overlap, TTR) and data independent (various distance measures extracted from the URIEL data base). Lauscher et al. (2020) study how well different similarity scores predict the success of the transfer and they find that language family is, in fact, the one that is least helpful in all the tasks considered (with mBERT and XLM-R). Various criteria for assessing language similarity remain an open research area in NLP (Turc et al., 2021; Pelloni et al., 2022; Samardžić et al., 2022; de Vries et al., 2022). Our proposal for assessing linguist diversity is relevant to these efforts too, as its key component is language comparison at the level of features extracted from both typological data bases and text samples.
+
+More generally, our work is intended to contribute to several wide-scope initiatives for improving the quality of data management in multilingual NLP (Bender and Friedman, 2018; Kreutzer et al., 2021; Lhoest et al., 2021) by focusing specifically on diversity assessments and data-independent scores for language comparison.
+
+# 3 Comparing Data Sets with Jaccard Similarity
+
+Our goal is to estimate the linguistic diversity of a data set with respect to some reference. Our score is thus a comparison between two data sets. More precisely, we compare scaled distributions
+
+![](images/5c44663fd5263138d2cc361837a03ac53382f7ae4db8776ed7f1b2ece09141bf.jpg)
+
+![](images/485964166e77c9c029e9e78008cce107d4168c89fe49bb96e740f4df28fa3e8f.jpg)
+
+![](images/072f5981cd83cb0581fea1cbe784ee622b81eb71488e4057aa66fd414821e838.jpg)  
+Figure 2: A toy example of comparing sets of measures with the minmax version of the Jaccard index.
+
+of the values of a numerical attribute as shown in Figure 2. The upper part of the figure shows (constructed) examples of two data sets (A and B), which we compare assuming that A is the data set whose diversity we want to assess and B is the reference. The values of the numerical attribute (one measurement per language) are on the x-axis and the numbers of languages are on the y-axis. Each bar in the figures represents the number of languages in the given data set with the numerical value in the given range (bin). For instance, the first bar in the upper left plot shows that the first sample (A) has 30 languages, with the values of their numerical attributes falling between 1 and 2. The other sample (B) has no languages in this bin.
+
+The width of the bins is arbitrary, but it does impact the score. Narrower bins capture more differences between two distributions than wider bins. By setting the width of the bins, we thus control the resolution at which we want to compare two data sets. In our example, the width is the distance between integers, but one can define other values (as long as the bins are of the same width).
+
+Since the data sets that we compare contain different numbers of languages, the values on the y-axis (counts of languages) are normalised in order to neutralise the effect of the size of the samples and focus rather on the diversity. We multiply all counts in the smaller set with the scalar  $c$ :
+
+$$
+c = \frac {\operatorname {m a x} (| A | , | B |)}{\operatorname {m i n} (| A | , | B |)} \tag {1}
+$$
+
+In this way, we increase the counts in the smaller set proportionally to obtain the same number of data points in both distributions and comparable numbers in each bin.
+
+Once we have represented our two sets in this way, we compare them using a generalised version of Jaccard similarity. This score shows how much the two distributions overlap. The original Jaccard index (Jaccard, 1912) compares two sets, but its generalised versions are suitable for comparing sets of measurements. Thus, we use the minmax version of the score  $(J_{mm})$ , initially proposed by Tanimoto (1958) for comparing vectors of binary values and then generalised to weight vectors by Grefenstette (1994). In our version, we compare two data sets as two vectors of weights: each bin is one dimension in the vectors and the number of languages in that bin is its weight.
+
+Intuitively, the score is the ratio between the intersection and the union of the two distributions (shown in the bottom part of Figure 2). Formally, we first map all the languages in all data sets to real numbers  $m:\mathbb{L}\mapsto \mathbb{R}$ , so that  $\{Y = m(x):x\in X\} = \{(x_i,y_i)\}$ , where  $x$  is a language in a data set,  $y$  is its corresponding measurement  $(y\in \mathbb{R})$  and the range of the index  $i1\dots |X|$  is the set of languages included in a data set. We then group the measurements into bins by applying a given threshold:  $\{Z = t(y):y\in Y\} = \{(y_i,z_j)\}$ , where  $z$  is the bin to which the measurement is assigned, the range of  $i$  is  $1\dots |X|$  and the range of  $j$  is  $1\dots |Z|$ .
+
+With this formalisation, we define the Jaccard minmax similarity of two data sets,  $J_{mm}(A,B)$ , as a similarity between two vectors of weights:
+
+$$
+J _ {m m} (\mathbf {a}, \mathbf {b}) = \frac {\sum_ {j = 1} ^ {| Z |} \operatorname* {m i n} \left(a _ {j} , b _ {j}\right)}{\sum_ {j = 1} ^ {| Z |} \operatorname* {m a x} \left(a _ {j} , b _ {j}\right)} \tag {2}
+$$
+
+The sum in the numerator represents the intersection and the sum in the denominator the union of the two sets of measurements. The weights  $a$  and  $b$  represent the number of measurements in the bin  $j$ . The values of  $J_{mm}$  fall in the range [0, 1], with higher values indicating more similarity between A and B, and, indirectly, better coverage of linguistic diversity in A.
+
+What is especially interesting about using  $J_{mm}$  as a diversity score is its transparency in terms of individual measurements: we can visualise and interpret where exactly a data set departs from the reference.
+
+# 4 Language Features
+
+We now turn to the question of how to define and take measures (the values on the x-axis in Figure 2) that can be used for calculating Jaccard minmax similarity between sets of languages. We use two kinds of descriptions.
+
+# 4.1 Grammar Features
+
+Typological data bases are currently the principal source of information about the properties of languages, but NLP researchers are faced with many obstacles when using this information. The popular software package lang2vec associated with the URIEL data base (Littell et al., 2017) alleviates some of the obstacles. First, the package solves the problem of incompatible feature values across different sources by mapping the data from several original data bases to binary features. Second, the problem of sparsity of feature values is solved by imputing the missing values: instead of a missing feature value in a language, the package returns the observed value for the same feature in the closest language. In this way, features become available for all queried languages, which is necessary for estimating language diversity, but a large proportion (roughly  $40\%$ ) of the returned features are imputed.
+
+While lang2vec facilitates retrieving typological features, its use for describing languages is limited due to remaining obstacles that are hard to solve. First, it does not contain any morphological features, which are especially relevant to NLP due to the known difficulties with that morphologically rich languages (Tsarfaty et al., 2013). The second unsolved problem is the fact that typological features are hard to add for languages for which they are not already available. Adding new features requires human expertise in many languages.
+
+# 4.2 Text Features
+
+As a complement to commonly used features from lang2vec, we make use of linguistically relevant text statistics. In this study, we focus on the mean word length as an approximation of aggregated morphological features, but other text-based features might be envisioned in future work. Our choice to start with word length relies on the observation that longer words can be expected in languages with rich morphology (large morphological paradigms, productive derivation), while shorter words are found in languages with less morphol-
+
+ogy. $^{4}$  As an empirical confirmation of the relationship between the word length and morphology, we perform a correlation test between the mean word length and morphological complexity calculated over morphological features (see Section 5 for the methods and Section 7 for the discussion).
+
+Text features are especially interesting in the context of NLP because they can be calculated automatically and applied to any language in which there are any texts to process. An important advantage of word length over other text statistics in this regard is that it manifests itself in very small samples of text and remains stable across different sizes. A sample of contiguous text of only 500 tokens gives us already a very good estimation of the overall mean word length. This can be seen in Figure 4 in the Appendix A, which shows the values of the mean word length on random samples of the length 500, 2000 and 10000 tokens in 87 languages. A correlation score (also in the Appendix A) shows that languages are almost identically ranked with all the sample sizes.
+
+# 4.3 Maximising Linguistic Diversity
+
+The editors of the WALS data base have selected two samples of languages (100 and 200 sample) as a means of guidance in the collective effort to create linguistic descriptions on a wide scale. These samples maximise genealogical (language family) and areal (geographic) diversity. Completing their descriptions is expected to minimise a potential bias regarding the relative frequency of different types of linguistic features included in the data base (Comrie et al., 2013). Figure 1 shows the locations of the languages in the 100 sample and their endangerment status according to UNESCO.
+
+Recently, text samples have been collected for most of the 100 languages in the TeDDi data set (Moran et al., 2022). These text data are sampled from online resources, e.g., Project Gutenberg, Open Subtitles (Lison and Tiedemann, 2016), The Parallel Bible Corpus (Mayer and Cysouw, 2014), the Universal Declaration of Human Rights, but also from grammars and other language documentation sources. For languages not present in online resources, the texts were manually transcribed.
+
+We take these two resources as the current reference that maximises linguistic diversity in terms of grammar features (WALS) and text features (TeDDi). We compare NLP data sets with these references, but our method can be applied to compare any given pair of data sets including potentially better references in the future.
+
+# 5 Data and Methods
+
+We calculate the Jaccard minmax diversity score  $(\mathrm{J}_{mm})$  for a number of popular multilingual data sets in comparison to the TeDDi sample. Without attempting to provide an exhaustive evaluation, we review data sets that satisfy the following criteria: multilingual (containing ten or more languages), relatively widely used and recently released or updated. The list is given in Table 1 and discussed in more detail in Section 6. For reference, we compare our  $\mathrm{J}_{mm}$  score to the typological index (TI) previously proposed as a linguistic diversity measure by Ponti et al. (2020) (see Section 2).
+
+Descriptions of the data sets often do not include all the information that was needed for our comparison. In particular, the number of language families is often not stated. To add this information, we extracted language names from the data files, converted these names into ISO 639-3 codes manually, and then retrieved the corresponding families from the Glottolog data base (top level family). Note that the conversion to ISO 639-3 codes led to some changes in the number of languages, compared to those cited in the data descriptions. For instance, the mBERT training data has only 97 distinct languages, not 104 as mentioned in the original description.
+
+# 5.1 Methods for Text Features
+
+We define words to be sequences of Unicode characters, delimited by spaces or other language-specific word delimiters, as defined by common multilingual tokenisers. We tokenise all the collected samples into word-level tokens using the Python library Polyglot (Al-Rfou, 2015). If a resulting token does not contain any alphanumeric characters, we discard it as punctuation. All the remaining tokens are further segmented into characters using the Python library segments (Moran and Cysouw, 2018). We split words into sequences of
+
+characters and take their length as word length. $^{11}$  We apply this same definition to all scripts, but we discuss below potential adjustments in the case of (partially) logographic scripts.
+
+Since the mean word length can be calculated on small samples, we take a single random sample for each language in a data set that we consider. To do this, we select a random position in the data set and extract contiguous text of the length up to 10K tokens starting from the random position. In case a data set does not contain such long texts (or sequences of paragraphs), we take smaller samples. The smallest samples are 200-300 tokens long.
+
+The output of these text processing steps is a set of real numbers, each number representing a language in a data set. To turn these numbers into discrete features, we group them into bins of equal size. We set the bin width to 1.[12]
+
+Mean word length vs. WALS features Following Bentz et al. (2016), we calculate a complexity score  $(C_{WALS})$  for each language using the set of 26 features that are relevant to describing morphology. This score is obtained by: 1) transforming the range of values each feature can take so that bigger values reflect the increasing use of morphology; 2) normalizing and averaging the resulting feature values per language. The list of features and transformations is given in Table 4 in Appendix B.  $C_{WALS}$  ranges from 0 to 1, where values closer to one indicate that the language encodes more morphosyntactic distinctions, making its morphology richer. All the values of the mean word length and morphological complexity for 29 diverse languages (the subset of TeDDi languages for which the 26 WALS features are known) are shown in Table 3 in Appendix B. We observe a strong correlation  $(\rho = 0.69)$ , which means that the variables quantify very similar phenomena and that the mean word length is a reasonable approximation of morphological types of languages. We return to this point in Section 7.
+
+Adjustments for logographic scripts Words in languages with logographic scripts tend to be shorter due to the fact that a single symbol corresponds to several alphabetic symbols (Sproat and Gutkin, 2021). For instance, in Mandarin Chinese, types such as the de (possessive particle), the le (aspect particle), is shi (copular verb 'is'), our wǒ-
+
+men (pronoun 'us') are assigned lengths (1, 1, 1, 2) respectively when measured in UTF-8 characters in the original script. When transliterated into Pinyin, the corresponding lengths are (2, 2, 3, 5). Hence, compared to Pinyin, the lengths are somewhat underestimated. It might seem more appropriate to convert the logographic scripts into their romanised counterparts to achieve cross-linguistic comparability. We opt for leaving such scripts without conversion, because we consider this phenomenon part of the diversity that we want to capture. Additional motivation for our choice is the fact that NLP systems have to deal with text as it is regardless of the mapping between written characters and sounds.
+
+Three languages in our data samples, Chinese, Japanese and Korean, are affected by this issue to a varied degree. In these cases, we scale the observed word length proportionally to the difference between the Chinese original script and Pinyin so that the scaled length is comparable to alphabetic scripts. Table 5 in Appendix C shows revised diversity scores after the adjustments.
+
+# 5.2 Linguistic Diversity Scores
+
+With the grammar features extracted from URIEL, we calculate syntactic diversity according to both TI and  $\mathbf{J}_{mm}$ .
+
+Syntax Typological Index  $(\mathbf{TI}_{syn})$  Following the formulation by Ponti et al. (2020), we calculate the typological index for each data set. In this context, a language is characterized by 103 syntactic features with binary values<sup>13</sup>. For each feature, Shannon entropy is estimated using the distribution of feature values in a data set. The feature-specific entropy values are averaged over the full set of features to obtain a TI score ranging from 0 to 1. The TI values closer to 1 indicate more diversity.
+
+Syntax Jaccard (J_mm_syn) We apply Jaccard similarity for comparing each data set against the TeDDi sample. Here the measures are the counts of the observed values of the same 103 syntactic feature available in 1ang2vec. This means that the items on the x-axis in Figure 2 are the 103 values, while the y-axis represents the number of times each feature value was observed in a data set. Since these feature values are binary, the width of the bin is not arbitrary in this case; it is determined by the values. Conceptually, grouping several features
+
+<table><tr><td>Name and main references</td><td>N(L)</td><td>N(F)</td><td>TI_syn</td><td>Jmm_syn</td><td>TI_morph</td><td>Jmm_morph</td></tr><tr><td>Universal Dependencies (UD)</td><td>106*</td><td>20*</td><td>0.567</td><td>0.736</td><td>0.349</td><td>0.650</td></tr><tr><td>Bible 100</td><td>103*</td><td>30*</td><td>0.649</td><td>0.811</td><td>0.311</td><td>0.534</td></tr><tr><td>mBERT</td><td>97*</td><td>15*</td><td>0.559</td><td>0.710</td><td>0.323</td><td>0.603</td></tr><tr><td>XTREME</td><td>40</td><td>14</td><td>0.612</td><td>0.775</td><td>0.311</td><td>0.457</td></tr><tr><td>XGLUE</td><td>19</td><td>7*</td><td>0.517</td><td>0.674</td><td>0.307</td><td>0.504</td></tr><tr><td>XNLI</td><td>15</td><td>7*</td><td>0.557</td><td>0.711</td><td>0.339</td><td>0.598</td></tr><tr><td>XCOPA</td><td>11</td><td>11</td><td>0.586</td><td>0.737</td><td>0.361</td><td>0.608</td></tr><tr><td>TyDiQA</td><td>11</td><td>10</td><td>0.626</td><td>0.751</td><td>0.343</td><td>0.525</td></tr><tr><td>XQuAD</td><td>12*</td><td>6*</td><td>0.523</td><td>0.680</td><td>0.341</td><td>0.588</td></tr><tr><td>TeDDi</td><td>89</td><td>51</td><td>0.706</td><td>-</td><td>0.369</td><td>-</td></tr></table>
+
+Table 1: Diversity of multilingual NLP data sets. The numbers in the second and the third column marked with an asterisk are added or modified by us. The numbers without an asterisk are reported in the respective publications. N(L): the number of languages in the data set. N(F): the number of families to which the languages belong. TI: typology index Ponti et al. (2020).  $J_{mm}$ : Jaccard minmax similarity (this paper).
+
+into a single one would correspond to increasing the bin width, but it is not clear at the moment how the features could be grouped. We thus work with the original set without any changes.
+
+With text features (mean word length) extracted from TeDDI and the scored NLP data sets, we calculate morphological diversity according to both TI and  $\mathrm{J}_{mm}$ .
+
+Morphology Typological index  $\left(\mathbf{TI}_{\mathit{morph}}\right)$  We adapt the measure proposed by Ponti et al. (2020) to the text-based features (mean word length). Each bin of the mean word length values is a feature and the number of languages that fall in a given bin are the counts of feature values. In other words, the mean word length becomes a vector of binary values, 1 for the languages that are in the bin and 0 for all the other languages in the sample. The rest of the calculation is the same as in  $\mathrm{TI}_{\mathrm{syn}}$ .
+
+Morphology Jaccard (J_mm_morph) Similarly to  $J_{-}mm_{-}\mathrm{syn}$ , we calculate the Jaccard score by comparing the distributions of the mean word length: TeDDi vs. a given NLP data set.
+
+# 6 Findings
+
+Table 1 lists all the reviewed data sets with all the measures of linguistic diversity. The colour scale of the cells represents the relative ranking of data sets according to each measure separately. TeDDI data set obtains the highest diversity scores at both levels (syntax and morphology) using the TI measure. This confirms the role of these resources as the current reference regarding linguistic diversity.
+
+TI and  $\mathbf{J}_{mm}$  are consistent The rankings of data sets according to the  $\mathbf{J}_{mm}$  score are very similar
+
+to those obtained with the TI score when the syntactic features are used. The agreement between the two measures is somewhat lower in the case of morphological features, but still rather high. The consistency between the two measures is not a trivial outcome given the entirely different approaches behind them. We can thus take this agreement as a validation of both measures. The main advantage of  $\mathrm{J}_{mm}$  compared to TI is its transparency regarding the kinds of languages that are missing. The difference with respect to the reference is visible at the level of features indicating the values that need to be added or removed to improve the diversity.
+
+Diversity rankings of NLP data sets The highest rankings appear split between the two structural levels. Bible 100 (Christodouloupoulos and Steedman, 2015) and XTREM (Hu et al., 2020) are the two most syntactically diverse data sets, while their morphological diversity is moderate to low. The Bible data set contains mostly non Indo-European languages, while the collection criteria for the XTREM data set was to maximise diversity. On the other hand, Universal Dependencies (UD, Nivre et al. (2020), which are often seen as especially biased towards European languages, show the best morphological, but a moderate syntactic diversity. XCOPA (Ponti et al., 2020) and TyDiQA (Clark et al., 2020) are data sets containing relatively few languages, but designed to maximise linguistic diversity. They are both highly ranked on  $3/4$  measures (two syntactic and one morphological). Contrary to this, the linguistic diversity ranking of one of the most popular benchmarks that contain manual labels for several downstream tasks, XGLUE (Liang et al., 2020; Wang et al., 2019) is
+
+![](images/aa791db69ee4a5869781157fef0028d7c50b61b5542c400dcfa42430f1229f87.jpg)
+
+![](images/b3d760642ba5bc28c41abcae5b726d3761766985d672fc626e045a99b8b167d6.jpg)
+
+![](images/7ce5b4f51df11ee765152472d271340ed9e46810fd2fae650336e8088db430ce.jpg)
+
+![](images/6b0343e5da51fe91d1f850d541582fbbbeceffdc75c430e5c922b609007b05b0.jpg)  
+Figure 3: Union and intersection between the distributions of the mean word length in TeDDi and NLP data sets.
+
+![](images/a635d77ac10849d7ce14fc95cbbe6f453e4efae11ec412413b519f6c05fe8883.jpg)
+
+![](images/192ccbdb7a0fcb090534cd9875c0e6f57289b937c3ed5f4b2222ff62bdbd875b.jpg)
+
+consistently low. XQuAD (Artetxe et al., 2020; Rajpurkar et al., 2016) fairs a little better, but it is still one of the least diverse data sets. The XNLI data set (Conneau et al., 2018; Bowman et al., 2015; Williams et al., 2018), which is compiled with the goal of spanning language families and which includes some low resource languages, remains of moderate linguistic diversity according to all measures. It is curious to see that the number of languages or even languages families included in a data set does not ensure a high linguistic diversity. For example, the mBERT $^{14}$  data set contains 97 languages in 15 language families, but it turns out to be less diverse than smaller data sets such as XCOPA (on  $\mathrm{TI}_{\mathrm{syn}}$ ,  $\mathrm{J}_{\mathrm{mm\_syn}}$  and  $\mathrm{J}_{\mathrm{mm\_morph}}$ ) and TyDiQA (on  $\mathrm{TI}_{\mathrm{syn}}$ ,  $\mathrm{J}_{\mathrm{mm\_syn}}$  and  $\mathrm{TI}_{\mathrm{morph}}$ ). The strategy of including the top 100 languages according to the size of their Wikipedia content (plus Thai and Mongolian), does not result in high diversity.
+
+Underrepresented language types Figure 3 is a visualisation of the  $J_{mm\_ morph}$  score<sup>15</sup> for some of the data sets showing the overlap and differences with the reference (TeDDi). The recurrent difference is whether a data set includes languages with long words or not (mean length  $>8$ ). Those that contain at least some languages with long words (UD, XCOPA) score much better on  $J_{mm\_ morph}$  than those that remain completely on
+
+the short-middle side (EXTREME, XGLUE, TyDiQA, mBERT). The second important factor that leads to lower scores is a strong peak of the distribution indicating a bias towards one of the length bins (EXTREME, XGLUE, mBERT). The third factor is a different ("wrong") shape of the distribution (TyDiQA). The data set that diverges the most is EXTREME, exhibiting all three factors of disagreement.
+
+The information about what kinds of languages are missing in a data set can be used to adjust language sampling and improve diversity. This is relatively straightforward when we deal with a single feature such as the mean word length. For example, the diversity of the mBERT language sample would be improved if the number of languages with a mean word length between 3 and 4 is reduced (by removing a given number of randomly selected languages). Instead of these languages, one should add a given number of languages with a mean word length greater than 7. It is not obvious where to look and how to find such languages (beyond the TeDDI sample), but knowing that they are needed might motivate such searches. Multifeature scores (such as feature entropy) could specify the needed languages more precisely, but they would require an optimisation method to ensure that a newly added language increases indeed the diversity score. It might happen, for instance, that we want to increase the count on one feature value but not on another. In this case, we need a language that has 1 on the desired feature value but 0 on the features that we do not want to change. Devising
+
+such a method is beyond the scope of the current paper, but it is a clear next step for future work.
+
+Overall, it seems that the right-hand side of the mean word length scale remains rather scarcely represented in all data sets, including the TeDDi sample itself. In future data collection, more effort should be put into representing languages with long words, especially because most of them are endangered. There are 12 languages in the TeDDI sample with a mean word length of over 7. If we localise them in Figure 1, we can see that ten of them are classified as extinct, endangered or vulnerable: Apurina (apu), Chukchi (ckt), Kalaallisut (kal), Kayardild (gyd), Makah (myh), Martuthunira (vma), Plains Cree (crk), Ngiyambaa (wyb), Wichita (wic) and Yagua (yad). Only two of these languages, Luvale (lue) and Zulu (zul) are safe.
+
+# 7 Discussion
+
+Our linguistic diversity scores include two kinds of language features (expert features extracted from data bases and the mean word length as a text feature) describing two structural levels (syntax and morphology). Readers not familiar with the details of how expert features are used in NLP might be left wondering whether the use of the mean word length is necessary and whether this measure is a good approximation of morphological types.
+
+Describing the use of expert features in NLP in Section 4, we note that the library lang2vec does not contain any morphological features, although these features are present in linguistic data bases. It is not clear why this is the case, but this means that morphological features are currently not used in NLP to assess linguistic diversity and the distances between languages. One possible reason for omitting morphological features could be the problem of sparsity, which would become even worse with these features leading to even more imputed values. For instance, if we want to study the distribution of 27 morphological features, only 34 languages will have a value for all these features. The values for the thousands of other languages would need to be imputed. This is the main reason why we propose to complement the existing sources of expert features with the mean word length as a value that can be easily calculated for any language on a small sample of text (500 tokens).
+
+To justify this proposal, we show that an independent measure of morphological complexity  $(C_{WALS})$  and the mean word length are strongly
+
+correlated, but the score of 0.69 means that the agreement is not perfect. A closer look into these two variables (Table 3 in Appendix B) points to the limitations of both measures, especially concerning the high values. For example, Turkish is the most complex language according to  $C_{WALS}$ , but its mean word length is well under 7. Although the correlation score is high and not due to chance, such aggregate measures remain approximations of the structural properties of languages. Nevertheless, these approximations are useful for tracking and improving linguistic diversity in data sets at the level of precision that is currently possible. Better approximations are certainly achievable in future work. Since our methods are general and can be applied to any set of features, any future improvements in representing linguistic structures can be easily integrated.
+
+# 8 Conclusion
+
+We have shown that the linguistic diversity of NLP data sets can be consistently assessed by two independent measures, TI (proposed in previous work) and  $\mathrm{J}_{mm}$  (proposed in this paper). Both of these measures show that a high number of languages and language families included in a data set is not sufficient to ensure linguistic diversity.
+
+To make the assessment of linguistic diversity automatic and rather simple, we show that text-based features such as the mean word length can be used as linguistic descriptors. These features can be easily calculated on very small text samples (of length of 500 tokens), overcoming the obstacles posed by the need to extract linguistic features from typological databases.
+
+An advantage of the  $\mathrm{J}_{mm}$  score over TI and other previous indicators of linguistic diversity is its capacity to show what kinds of languages are missing in a given data set in comparison to a reference. Assessing popular NLP data sets with this measure revealed that the most underrepresented languages are those with rich morphology. This kind of direct and transparent comparison can improve multilingual NLP coverage in the long run.
+
+# Acknowledgements
+
+This research is partially supported by the Swiss National Science Foundation (SNSF) grants 176305 and PCEFP1_186841. We thank the anonymous reviewers for their suggestions, which have improved the clarity of the paper.
+
+# Limitations
+
+A limitation of our study is that the two levels of linguistic structures are represented with different features: syntax with expert features from linguistic data bases and morphology with mean word length as a text feature. Our results suggest that the two measures agree more at the level of syntax than at the level of morphology. To draw sound conclusions about the impact of the structural level on the agreement between the two measures, we would need both kinds of features for both levels. While we indirectly compare text and expert features at the level of morphology (via the correlation test), we do not propose syntactic features that could be extracted from text. We focused here on the current gap in the available linguistic features (the lack of morphological features in lang2vec), but devising and validating text-based syntactic features would deserve more attention in future work.
+
+# References
+
+Rami Al-Rfou. 2015. *Polyglot: A massive multilingual natural language processing pipeline*. Ph.D. thesis, State University of New York at Stony Brook.  
+Mikel Artetxe, Sebastian Ruder, and Dani Yogatama. 2020. On the cross-lingual transferability of monolingual representations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4623–4637, Online. Association for Computational Linguistics.  
+Emily M. Bender and Batya Friedman. 2018. Data statements for natural language processing: Toward mitigating system bias and enabling better science. Transactions of the Association for Computational Linguistics, 6:587-604.  
+Christian Bentz, Tatyana Ruzsics, Alexander Koplenig, and Tanja Samardzic. 2016. A comparison between morphological complexity measures: typological data vs. language corpora. In Proceedings of the workshop on computational linguistics for linguistic complexity (cl4lc), pages 142-153.  
+Aleksandrs Berdicevskis, Căgni Čöltekin, Katharina Ehret, Kilu von Prince, Daniel Ross, Bill Thompson, Chunxiao Yan, Vera Demberg, Gary Lupyan, Taraka Rama, et al. 2018. Using universal dependencies in cross-linguistic complexity research. In Proceedings of the Second Workshop on Universal Dependencies (UDW 2018), pages 8-17.  
+Balthasar Bickel, Johanna Nichols, Taras Zakharko, Alena Witzlack-Makarevich, Kristine Hildebrandt, Michael Rießler, Lennart Bierkandt, Fernando Zúñiga, and John B Lowe. 2017. The autotyp typological databases. version 0.1.2.
+
+Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. 2015. A large annotated corpus for learning natural language inference. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 632-642, Lisbon, Portugal. Association for Computational Linguistics.  
+Christos Christodouloupoulos and Mark Steedman. 2015. A massively parallel corpus: the bible in 100 languages. Language Resources and Evaluation, 49(2):375-395.  
+Jonathan H. Clark, Eunsol Choi, Michael Collins, Dan Garrette, Tom Kwiatkowski, Vitaly Nikolaev, and Jennimaria Palomaki. 2020. TyDi QA: A benchmark for information-seeking question answering in typologically diverse languages. Transactions of the Association for Computational Linguistics, 8:454-470.  
+Bernard Comrie, Matthew S. Dryer, David Gil, and Martin Haspelmath. 2013. Introduction. In Matthew S. Dryer and Martin Haspelmath, editors, The World Atlas of Language Structures Online. Max Planck Institute for Evolutionary Anthropology, Leipzig.  
+Alexis Conneau, Rudy Rinott, Guillaume Lample, Adina Williams, Samuel Bowman, Holger Schwenk, and Veselin Stoyanov. 2018. XNLI: Evaluating cross-lingual sentence representations. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2475-2485, Brussels, Belgium. Association for Computational Linguistics.  
+Wietse de Vries, Martijn Wieling, and Malvina Nissim. 2022. Make the best of cross-lingual transfer: Evidence from POS tagging with over 100 languages. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7676-7685, Dublin, Ireland. Association for Computational Linguistics.  
+Matthew S. Dryer and Martin Haspelmath, editors. 2013. WALS Online. Max Planck Institute for Evolutionary Anthropology, Leipzig.  
+Gregory Grefenstette. 1994. Explorations in Automatic Thesaurus Discovery. Kluwer Academic Publishers, USA.  
+Harald Hammarström, Robert Forkel, and Martin Haspelmath. 2018. Glottolog 3.3. Leipzig.  
+Martin Haspelmath. 2007. Pre-established categories don't exist: Consequences for language description and typology. Linguistic Typology, 11(1):119-132.  
+Junjie Hu, Sebastian Ruder, Aditya Siddhant, Graham Neubig, Orhan First, and Melvin Johnson. 2020. XTREME: A massively multilingual multitask benchmark for evaluating cross-lingual generalisation. In Proceedings of the 37th International
+
+Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, volume 119 of Proceedings of Machine Learning Research, pages 4411-4421. PMLR.  
+P. Jaccard. 1912. The distribution of the flora in the alpine zone.1. New Phytologist, 11:37-50.  
+Pratik Joshi, Sebastin Santy, Amar Budhiraja, Kalika Bali, and Monojit Choudhury. 2020. The state and fate of linguistic diversity and inclusion in the NLP world. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6282-6293, Online. Association for Computational Linguistics.  
+Kimmo Kettunen. 2014. Can type-token ratio be used to show morphological complexity of languages? Journal of Quantitative Linguistics, 21(3):223-245.  
+Julia Kreutzer, Isaac Caswell, Lisa Wang, Ahsan Wahab, Daan van Esch, Nasanbayar Ulzii-Orshikh, Alahsera Tapo, Nishant Subramani, Artem Sokolov, Claytone Sikasote, Monang Setyawan, Supheakmungkol Sarin, Sokhar Samb, Benoit Sagot, Clara Rivera, Annette Rios, Isabel Papadimitriou, Salomey Osei, Pedro Ortiz Suárez, Iroro Orife, Kelechi Ogueji, Andre Niyongabo Rubungo, Toan Q. Nguyen, Mathias Müller, André Müller, Shamsuddeen Hassan Muhammad, Nanda Muhammad, Ayanda Mnyakeni, Jamshidbek Mirzakhalov, Tapiwanashe Matangira, Colin Leong, Nze Lawson, Sneha Kudugunta, Yacine Jernite, Mathias Jenny, Orhan First, Bonaventure F. P. Dossou, Sakhile Dlamini, Nisansa de Silva, Sakine Cabuk Balli, Stella Biderman, Alessia Battisti, Ahmed Baruwa, Ankur Bapna, Pallavi Baljekar, Israel Abebe Azime, Ayodele Awokoya, Duygu Ataman, Orevaoghene Ahia, Oghenefego Ahia, Sweta Agrawal, and Mofetoluwa Adeyemi. 2021. Quality at a glance: An audit of web-crawled multilingual datasets.  
+Anne Lauscher, Vinit Ravishankar, Ivan Vulic, and Goran Glavaš. 2020. From zero to hero: On the limitations of zero-shot language transfer with multilingual Transformers. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4483-4499, Online. Association for Computational Linguistics.  
+Quentin Lhoest, Albert Villanova del Moral, Yacine Jernite, Abhishek Thakur, Patrick von Platen, Suraj Patil, Julien Chaumont, Mariama Drame, Julien Plu, Lewis Tunstall, Joe Davison, Mario Šaško, Gunjan Chhablani, Bhavitvya Malik, Simon Brandeis, Teven Le Scao, Victor Sanh, Canwen Xu, Nicolas Patry, Angelina McMillan-Major, Philipp Schmid, Sylvain Gugger, Clément Delangue, Theo Matussière, Lysandre Debut, Stas Bekman, Pierrick Cistac, Thibault Goehringer, Victor Mustar, François Lagunas, Alexander Rush, and Thomas Wolf. 2021. Datasets: A community library for natural language processing. In Proceedings of the 2021 Conference
+
+on Empirical Methods in Natural Language Processing: System Demonstrations, pages 175-184, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Yaobo Liang, Nan Duan, Yeyun Gong, Ning Wu, Fenfei Guo, Weizhen Qi, Ming Gong, Linjun Shou, Daxin Jiang, Guihong Cao, Xiaodong Fan, Ruofei Zhang, Rahul Agrawal, Edward Cui, Sining Wei, Taroon Bharti, Ying Qiao, Jiun-Hung Chen, Winnie Wu, Shuguang Liu, Fan Yang, Daniel Campos, Rangan Majumder, and Ming Zhou. 2020. XGLUE: A new benchmark dataset for cross-lingual pretraining, understanding and generation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6008-6018, Online. Association for Computational Linguistics.  
+Yu-Hsiang Lin, Chian-Yu Chen, Jean Lee, Zirui Li, Yuyan Zhang, Mengzhou Xia, Shruti Rijhwani, Junxian He, Zhisong Zhang, Xuezhe Ma, Antonios Anastasopoulos, Patrick Littell, and Graham Neubig. 2019. Choosing transfer languages for cross-lingual learning. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3125-3135, Florence, Italy. Association for Computational Linguistics.  
+Pierre Lison and Jörg Tiedemann. 2016. Opensubtitles2016: Extracting large parallel corpora from movie and tv subtitles. In Proceedings from LREC 2016, pages 923-929. European Language Resources Association.  
+Patrick Littell, David R. Mortensen, Ke Lin, Katherine Kairis, Carlisle Turner, and Lori Levin. 2017. URIEL and lang2vec: Representing languages as typological, geographical, and phylogenetic vectors. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, pages 8-14, Valencia, Spain. Association for Computational Linguistics.  
+Thomas Mayer and Michael Cysouw. 2014. Creating a massively parallel bible corpus. In Proceedings of the International Conference on Language Resources and Evaluation (LREC), pages 3158-3163.  
+Steven Moran. 2016. The ACQDIV database: Min(d)ing the ambient language. In Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC'16), pages 4423-4429, Portoorž, Slovenia. European Language Resources Association (ELRA).  
+Steven Moran, Christian Bentz, Ximena Gutierrez-Vasques, Olga Pelloni, and Tanja Samardzic. 2022. TeDDi sample: Text data diversity sample for language comparison and multilingual NLP. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 1150-1158, Marseille, France. European Language Resources Association.
+
+Steven Moran and Michael Cysouw. 2018. The Unicode cookbook for linguists. Number 10 in Translation and Multilingual Natural Language Processing. Language Science Press, Berlin.  
+Joakim Nivre, Marie-Catherine de Marneffe, Filip Ginter, Jan Hajic, Christopher D. Manning, Sampo Pyysalo, Sebastian Schuster, Francis Tyers, and Daniel Zeman. 2020. Universal Dependencies v2: An evergrowing multilingual treebank collection. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 4034-4043, Marseille, France. European Language Resources Association.  
+Olga Pelloni, Anastassia Shaitarova, and Tanja Samardzic. 2022. Subword evenness (SuE) as a predictor of cross-lingual transfer to low-resource languages. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 7428-7445, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Telmo Pires, Eva Schlinger, and Dan Garrette. 2019. How multilingual is multilingual BERT? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4996-5001, Florence, Italy. Association for Computational Linguistics.  
+Edoardo Maria Ponti, Goran Glavaš, Olga Majewska, Qianchu Liu, Ivan Vulić, and Anna Korhonen. 2020. XCOPA: A multilingual dataset for causal commonsense reasoning. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2362-2376, Online. Association for Computational Linguistics.  
+Edoardo Maria Ponti, Helen O'Horan, Yevgeni Berzak, Ivan Vulic, Roi Reichart, Thierry Poibeau, Ekaterina Shutova, and Anna Korhonen. 2019. Modeling language variation and universals: A survey on typological linguistics for natural language processing. Computational Linguistics, 45(3):559-601.  
+Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ questions for machine comprehension of text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383-2392, Austin, Texas. Association for Computational Linguistics.  
+Tanja Samardžić, Ximena Gutierrez-Vasques, Rob van der Goot, Max Müller-Eberstein, Olga Pelloni, and Barbara Plank. 2022. On language spaces, scales and cross-lingual transfer of UD parsers. In Proceedings of the 26th Conference on Computational Natural Language Learning (CoNLL), pages 266–281, Abu Dhabi, United Arab Emirates (Hybrid). Association for Computational Linguistics.  
+Richard Sproat and Alexander Gutkin. 2021. The Taxonomy of Writing Systems: How to Measure How Logographic a System Is. Computational Linguistics, 47(3):477-528.
+
+Sabine Stoll and Balthasar Bickel. 2013. Capturing diversity in language acquisition research. In Balthasar Bickel, Lenore A. Grenoble, David A. Peterson, and Alan Timberlake, editors, Language typology and historical contingency: studies in honor of Johanna Nichols, pages 195-260. Benjamins, Amsterdam. [pre-print available at http://www.psycholinguistics.uzh.ch/stoll/publications/stollbickel.sampling2012rev.pdf].  
+T. T Tanimoto. 1958. An elementary mathematical theory of classification and prediction. International Business Machines Corporation.  
+Reut Tsarfaty, Djamé Seddah, Sandra Kübler, and Joakim Nivre. 2013. Parsing morphologically rich languages: Introduction to the special issue. Computational Linguistics, 39(1):15-22.  
+Iulia Turc, Kenton Lee, Jacob Eisenstein, Ming-Wei Chang, and Kristina Toutanova. 2021. Revisiting the primacy of english in zero-shot cross-lingual transfer.  
+Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. 2019. GLUE: A multi-task benchmark and analysis platform for natural language understanding. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net.  
+Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. A broad-coverage challenge corpus for sentence understanding through inference. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1112-1122, New Orleans, Louisiana. Association for Computational Linguistics.
+
+# A Mean Word Length Correlation between Different Sample Size
+
+![](images/154462964f9a90dc5a5d1ee6832b4508d0a7d3437df4a8d0954cb1ba47c0d136.jpg)  
+Figure 4: Mean word length measures at different text sizes in TeDDi. The languages on the x-axis are sorted according to the increasing value calculated on the biggest sample (10K). The values in the two smaller samples (2K and 500) depart very little from the main trend.
+
+To make sure that the stability across different sample sizes suggested by Figure 4 is not a mere consequence of a relatively small range of variation, we perform correlation tests between different samples and in comparison to other measures (TTR and unigram entropy (H)). Table 2 shows that the ranks of languages change considerably less across different sample sizes when considering the mean word length than in the other two measures.
+
+<table><tr><td>Samples</td><td>MWL</td><td>H</td><td>TTR</td></tr><tr><td>500 tokens vs. max.</td><td>0.99</td><td>0.85</td><td>0.84</td></tr><tr><td>2K tokens vs. max</td><td>0.99</td><td>0.95</td><td>0.94</td></tr></table>
+
+Table 2: Spearman rank correlation showing how much rankings of languages change with text measures taken on random samples of different size.
+
+B Word length and morphological complexity  
+
+<table><tr><td>ISO396-3</td><td>Name</td><td>MWL</td><td>CwALS</td></tr><tr><td>abk</td><td>Abkhazian</td><td>7.17</td><td>0.62</td></tr><tr><td>apu</td><td>Apurinã</td><td>7.67</td><td>0.60</td></tr><tr><td>arz</td><td>Egyptian Arabic</td><td>4.44</td><td>0.49</td></tr><tr><td>bsn</td><td>Barasana-Eduria</td><td>6.02</td><td>0.69</td></tr><tr><td>ckt</td><td>Chukchi</td><td>8.45</td><td>0.50</td></tr><tr><td>deu</td><td>German</td><td>4.87</td><td>0.55</td></tr><tr><td>ell</td><td>Modern Greek</td><td>4.72</td><td>0.53</td></tr><tr><td>eng</td><td>English</td><td>4.18</td><td>0.42</td></tr><tr><td>eus</td><td>Basque</td><td>5.70</td><td>0.64</td></tr><tr><td>fin</td><td>Finnish</td><td>6.23</td><td>0.66</td></tr><tr><td>fra</td><td>French</td><td>4.41</td><td>0.45</td></tr><tr><td>hae</td><td>Eastern Oromo</td><td>5.91</td><td>0.53</td></tr><tr><td>hau</td><td>Hausa</td><td>4.08</td><td>0.38</td></tr><tr><td>heb</td><td>Modern Hebrew</td><td>3.94</td><td>0.54</td></tr><tr><td>ind</td><td>Indonesian</td><td>5.42</td><td>0.40</td></tr><tr><td>kan</td><td>Kannada</td><td>5.22</td><td>0.65</td></tr><tr><td>kat</td><td>Georgian</td><td>4.78</td><td>0.50</td></tr><tr><td>khk</td><td>Halh Mongolian</td><td>5.66</td><td>0.53</td></tr><tr><td>kut</td><td>Kutenai</td><td>4.60</td><td>0.37</td></tr><tr><td>lvk</td><td>Lavukaleve</td><td>4.77</td><td>0.67</td></tr><tr><td>qvi</td><td>Imbabura Highland Quichua</td><td>8.18</td><td>0.71</td></tr><tr><td>rus</td><td>Russian</td><td>4.79</td><td>0.52</td></tr><tr><td>spa</td><td>Spanish</td><td>4.37</td><td>0.45</td></tr><tr><td>swh</td><td>Swahili</td><td>5.72</td><td>0.71</td></tr><tr><td>tur</td><td>Turkish</td><td>6.07</td><td>0.76</td></tr><tr><td>vie</td><td>Vietnamese</td><td>3.20</td><td>0.21</td></tr><tr><td>yaq</td><td>Yaqui</td><td>5.31</td><td>0.57</td></tr><tr><td>yor</td><td>Yoruba</td><td>3.52</td><td>0.25</td></tr></table>
+
+Spearmann correlation  $\rho = 0.69$
+
+Table 3: Mean Word length (MWL) and morphological complexity measure ( $C_{WALS}$ ) in the subset of TeDDi languages for which 26 WALS morphology features are known.
+
+<table><tr><td>Chapter</td><td>Name</td><td>Categories</td><td>Transformation</td><td>Final Values</td></tr><tr><td>22A</td><td>Inflectional Synthesis</td><td>7 (ordinal)</td><td>none</td><td>1-7</td></tr><tr><td>26A</td><td>Prefixing vs. Suffixing in Inflectional Morphology</td><td>6 (non-ordinal)</td><td>binarization</td><td>0-1</td></tr><tr><td>27A</td><td>Reduplication</td><td>3 (non-ordinal)</td><td>binarization</td><td>0-1</td></tr><tr><td>28A</td><td>Case Syncretism</td><td>4 (ordinal)</td><td>reorder</td><td>1-4</td></tr><tr><td>29A</td><td>Syncretism in Verbal Per-son/Number marking</td><td>3 (ordinal)</td><td>none</td><td>1-3</td></tr><tr><td>30A</td><td>Number of Genders</td><td>5 (ordinal)</td><td>none</td><td>1-5</td></tr><tr><td>33A</td><td>Coding of Nominal Plurality</td><td>9 (partially ordinal)</td><td>binarization</td><td>0-1</td></tr><tr><td>34A</td><td>Occurrence of Nominal Plurality</td><td>6 (ordinal)</td><td>none</td><td>1-6</td></tr><tr><td>49A</td><td>Number of Cases</td><td>9 (ordinal)</td><td>remove</td><td>1-8</td></tr><tr><td>51A</td><td>Position of Case Affixes</td><td>9 (non-ordinal)</td><td>binarization</td><td>0-1</td></tr><tr><td>57A</td><td>Position of Pronominal Posses-sive Affixes</td><td>4 (non-ordinal)</td><td>binarization</td><td>0-1</td></tr><tr><td>59A</td><td>Possessive Classification</td><td>4 (ordinal)</td><td>none</td><td>1-4</td></tr><tr><td>65A</td><td>Perfective/Imperfective Aspect</td><td>binary</td><td>none</td><td>0-1</td></tr><tr><td>66A</td><td>The Past Tense</td><td>4 (ordinal)</td><td>reorder</td><td>1-4</td></tr><tr><td>67A</td><td>The Future Tense</td><td>binary</td><td>none</td><td>0-1</td></tr><tr><td>69A</td><td>Position of Tense/Aspect Affixes</td><td>5 (non-ordinal)</td><td>binarization</td><td>0-1</td></tr><tr><td>70A</td><td>The Morphological Imperative</td><td>5 (partially ordinal)</td><td>recategorization</td><td>1-4</td></tr><tr><td>73A</td><td>The Optative</td><td>binary</td><td>none</td><td>0-1</td></tr><tr><td>74A</td><td>Situational Possibility</td><td>3 (non-ordinal)</td><td>binarization</td><td>0-1</td></tr><tr><td>75A</td><td>Epistemic Possibility</td><td>3 (non-ordinal)</td><td>binarization</td><td>0-1</td></tr><tr><td>78A</td><td>Coding of Evidentiality</td><td>6 (non-ordinal)</td><td>binarization</td><td>0-1</td></tr><tr><td>94A</td><td>Subordination</td><td>5 (non-ordinal)</td><td>binarization</td><td>0-1</td></tr><tr><td>101A</td><td>Expression of Pronominal Sub-jects</td><td>6 (non-ordinal)</td><td>binarization</td><td>0-1</td></tr><tr><td>102A</td><td>Verbal Person Marking</td><td>5 (partially ordinal)</td><td>recategorization</td><td>1-3</td></tr><tr><td>111A</td><td>Nonperiphrastic Causative Constructions</td><td>4 (non-ordinal)</td><td>binarization</td><td>0-1</td></tr><tr><td>112A</td><td>Negative Morphemes</td><td>6 (non-ordinal)</td><td>binarization</td><td>0-1</td></tr></table>
+
+Table 4: Subset of WALS features that we use for characterizing the morphological complexity of languages. The column "Final Values" gives the range of values each feature can take after transformations were performed to the original values (Bentz et al., 2016)
+
+# C Word Length Adjustments for Logographic Scripts
+
+<table><tr><td>Name and main references</td><td>N(L)</td><td>N(F)</td><td>TI_syn</td><td>Jmm_syn</td><td>TI_morph</td><td>Jmm_morph</td></tr><tr><td>Universal Dependencies (UD)</td><td>106*</td><td>20*</td><td>0.567</td><td>0.736</td><td>0.337</td><td>0.665</td></tr><tr><td>Bible 100</td><td>103*</td><td>30*</td><td>0.649</td><td>0.811</td><td>0.302</td><td>0.617</td></tr><tr><td>mBERT</td><td>97*</td><td>15*</td><td>0.559</td><td>0.710</td><td>0.316</td><td>0.617</td></tr><tr><td>XTREME</td><td>40</td><td>14</td><td>0.612</td><td>0.775</td><td>0.311</td><td>0.471</td></tr><tr><td>XGLUE</td><td>19</td><td>7*</td><td>0.517</td><td>0.674</td><td>0.297</td><td>0.580</td></tr><tr><td>XNLI</td><td>15</td><td>7*</td><td>0.557</td><td>0.711</td><td>0.321</td><td>0.704</td></tr><tr><td>XCOPA</td><td>11</td><td>11</td><td>0.586</td><td>0.737</td><td>0.336</td><td>0.634</td></tr><tr><td>TyDiQA</td><td>11</td><td>10</td><td>0.626</td><td>0.751</td><td>0.343</td><td>0.552</td></tr><tr><td>XQuAD</td><td>12*</td><td>6*</td><td>0.523</td><td>0.680</td><td>0.318</td><td>0.634</td></tr><tr><td>TeDDi</td><td>89</td><td>51</td><td>0.706</td><td>-</td><td>0.361</td><td>-</td></tr></table>
+
+Table 5: Diversity of multilingual NLP data sets with adjustments for logographic scripts. Compared to the main results in Table 1, all  $\mathrm{TI}_{\mathrm{morph}}$  scores are slightly decreased and  $\mathbf{J}_{mm\_ morph}$  slightly increased. The rankings of the t are mostly preserved, with the exception of XNLI, whose  $\mathbf{J}_{mm\_ morph}$  ranking improves.

2024/A Measure for Transparent Comparison of Linguistic Diversity in Multilingual NLP Data Sets/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9808b31c14d1badb3adf0231c5705a1dcf04d01e63a4fa85a7dd959b1d9b7c58
+size 658042

2024/A Novel Paradigm Boosting Translation Capabilities of Large Language Models/8eb804a5-0f10-4fdb-af4a-6a552f7a4c4e_content_list.json

@@ -0,0 +1,1699 @@
+[
+    {
+        "type": "text",
+        "text": "A Novel Paradigm Boosting Translation Capabilities of Large Language Models",
+        "text_level": 1,
+        "bbox": [
+            126,
+            89,
+            870,
+            129
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Jiaxin Guo*, Hao Yang†, Zongyao Li",
+        "bbox": [
+            341,
+            147,
+            660,
+            164
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Daimeng Wei, Hengchao Shang, Xiaoyu Chen",
+        "bbox": [
+            302,
+            164,
+            699,
+            180
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{jiaxinguo1,yanghao30,lizongyao}  $@$  huawei.com",
+        "bbox": [
+            302,
+            181,
+            702,
+            197
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{weidaimeng,shanghengchao,chenxiaoyu35}@huawei.com",
+        "bbox": [
+            262,
+            198,
+            742,
+            212
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Huawei Translation Services Center, Beijing, China",
+        "bbox": [
+            292,
+            214,
+            710,
+            229
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            260,
+            339,
+            275
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "This paper presents a study on strategies to enhance the translation capabilities of large language models (LLMs) in the context of machine translation (MT) tasks. The paper proposes a novel paradigm consisting of three stages: Secondary Pre-training using Extensive Monolingual Data, Continual Pre-training with Interlinear Text Format Documents, and Leveraging Source-Language Consistent Instruction for Supervised Fine-Tuning. Previous research on LLMs focused on various strategies for supervised fine-tuning (SFT), but their effectiveness has been limited. While traditional machine translation approaches rely on vast amounts of parallel bilingual data, our paradigm highlights the importance of using smaller sets of high-quality bilingual data. We argue that the focus should be on augmenting LLMs' cross-lingual alignment abilities during pre-training rather than solely relying on extensive bilingual data during SFT. Experimental results conducted using the Llama2(Touvron et al., 2023) model, particularly on Chinese-Llama2(Cui et al., 2023) after monolingual augmentation, demonstrate the improved translation capabilities of LLMs. A significant contribution of our approach lies in Stage2: Continual Pre-training with Interlinear Text Format Documents, which requires less than 1B training data, making our method highly efficient. Additionally, in Stage3, we observed that setting instructions consistent with the source language benefits the supervised fine-tuning process. Experimental results demonstrate that our approach surpasses previous work and achieves superior performance compared to models such as NLLB-54B(Team et al., 2022) and GPT3.5-text-davinci-003, despite having a significantly smaller parameter count of only 7B or 13B. This achievement establishes our method as a pioneering strategy in the field of machine translation.",
+        "bbox": [
+            142,
+            288,
+            460,
+            888
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            510,
+            260,
+            653,
+            275
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Translation capabilities of large language models (LLMs)(Brown et al., 2020; Chowdhery et al., 2023; Touvron et al., 2023) in machine translation (MT) tasks have been explored extensively in previous research(Jiao et al., 2023; Zeng et al., 2023; Chen et al., 2023; Xu et al., 2023; Yang et al., 2023; Zhang et al., 2023). However, achieving significant improvements in translation quality through supervised fine-tuning (SFT) strategies has proven challenging. Traditionally, machine translation relies on vast amounts of parallel bilingual data, but SFT only requires a small amount of high-quality bilingual data, highlighting a crucial distinction. It is a naive approach to consider using vast quantities of parallel bilingual data during SFT. However, experiments have shown that increasing the data volume yields limited improvements in quality and even leads to performance degradation on certain test sets. Thus, the question arises: are extensive parallel bilingual data useless in SFT, or are they being misused?",
+        "bbox": [
+            507,
+            288,
+            885,
+            627
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we propose a novel training paradigm, consisting of three stages, to boost the translation capabilities of LLMs. Our contributions include refining the training strategy for downstream tasks and emphasizing the enhancement of LLMs' cross-lingual alignment abilities during pretraining. These contributions address the limitations observed in previous research. Our training paradigm comprises the following stages:",
+        "bbox": [
+            507,
+            629,
+            884,
+            772
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Stage 1: Continual Pre-training using Extensive Monolingual Data. Consistent with previous findings(Xu et al., 2023), we validate the effectiveness of monolingual data augmentation. Specifically, we perform SFT on a Chinese-Llama2(Cui et al., 2023) model, which undergoes monolingual data augmentation, thereby demonstrating improved translation capabilities.",
+        "bbox": [
+            507,
+            775,
+            884,
+            903
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Stage 2: Continual Pre-training with",
+        "bbox": [
+            526,
+            904,
+            880,
+            920
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Co-first Author",
+        "bbox": [
+            137,
+            894,
+            238,
+            904
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "†Co-First author",
+        "bbox": [
+            137,
+            906,
+            238,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "639",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: NAACL 2024, pages 639-649",
+        "bbox": [
+            226,
+            945,
+            769,
+            957
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "June 16-21, 2024 ©2024 Association for Computational Linguistics",
+        "bbox": [
+            292,
+            959,
+            705,
+            971
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Sentence-aligned Parallel Data. We construct interlinear text format from sentence-aligned bilingual parallel data and utilize them for continual pre-training of LLMs. Experimental results demonstrate the critical importance of this stage, resulting in a significant improvement in translation quality, particularly for English-Other translations. Stage 2 stands as a pivotal contribution in our paper, requiring less than 1B training data, thereby enhancing training efficiency.",
+        "bbox": [
+            112,
+            84,
+            489,
+            244
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Stage 3: Leveraging Source-Language Consistent Instruction for Supervised Fine-Tuning. In SFT, we discover that using instruction aligned with the source language of the translation notably improves performance. Leveraging source-language consistent instructions during SFT yields significant enhancements.",
+        "bbox": [
+            112,
+            246,
+            489,
+            357
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In summary, we introduce a three-stage training paradigm, highlighting the effectiveness of secondary pre-training, continual pre-training with interlinear text format documents, and leveraging source-language consistent instruction for supervised fine-tuning. These contributions address the limitations observed in previous research and pave the way for improved translation quality.",
+        "bbox": [
+            112,
+            360,
+            489,
+            488
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            112,
+            502,
+            268,
+            518
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1 Large Language Models",
+        "text_level": 1,
+        "bbox": [
+            112,
+            532,
+            352,
+            546
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Foundation Model Foundation Model, a product of pre-training, is a prominent type of Large Language Model. It has gained substantial recognition in recent years for its impressive capabilities in natural language processing tasks. The most prevalent architectural framework for such models is the Transformer, which employs a series of self-attention mechanisms to process input text efficiently.",
+        "bbox": [
+            112,
+            554,
+            489,
+            697
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Among the state-of-the-art Large Language Models, notable examples include GPT-3(Brown et al., 2020) and Llama2(Touvron et al., 2023). These models have been widely lauded for their exceptional proficiency in understanding and generating natural language text. They showcase the remarkable potential of Foundation Models, pushing the boundaries of language processing and setting new benchmarks in various applications.",
+        "bbox": [
+            112,
+            700,
+            489,
+            844
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Instruct/Chat Model Instruct/Chat Model, a variant of Large Language Models, is specifically developed through the process of Supervised Fine-Tuning (SFT). Unlike Foundation Models, which",
+        "bbox": [
+            112,
+            857,
+            489,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "are pre-trained, Instruct/Chat Models undergo additional supervised training to enhance their performance in specific tasks such as instruction following or conversational dialogue.",
+        "bbox": [
+            507,
+            84,
+            884,
+            148
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Supervised Fine-Tuning involves training the model on labeled datasets, where human annotators provide examples of desired input-output behavior. This approach enables Instruct/Chat Models to learn task-specific skills and exhibit improved performance in situations that require language understanding, generation, and interaction.",
+        "bbox": [
+            507,
+            149,
+            884,
+            261
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Noteworthy advancements have been observed in Instruct/Chat Models, with notable examples including models like ChatGPT. These models have exhibited remarkable outcomes in conversational scenarios, demonstrating their potential in enabling interactive and engaging human-like conversations.",
+        "bbox": [
+            507,
+            263,
+            884,
+            359
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2 Machine Translation Task",
+        "text_level": 1,
+        "bbox": [
+            507,
+            374,
+            761,
+            388
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Machine Translation Task refers to the process of automatically translating text from one language to another using computational methods.",
+        "bbox": [
+            507,
+            397,
+            882,
+            444
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Traditional Methods Traditional machine translation methods primarily rely on encoder-decoder(Vaswani et al., 2017) models, where an encoder converts the source language sentence and a decoder produces the translated sentence. These methods heavily depend on large bilingual parallel corpora for training, aligning source sentences with their corresponding translations. Data augmentation(Sennrich et al., 2016; Wei et al., 2023) is a common practice in traditional machine translation. Some studies(Gu et al., 2018; Ghazvininejad et al., 2019; Wang et al., 2021; Guo et al., 2021; Yu et al., 2021) also investigate transforming them into parallel architectures to speed up inference efficiency.",
+        "bbox": [
+            507,
+            458,
+            884,
+            682
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "LLM-based Methods In recent years, Language Model (LM)-based approaches have gained attention in the field of machine translation. These approaches leverage pre-trained language models, such as the GPT (Generative Pre-trained Transformer)(Brown et al., 2020; Chowdhery et al., 2023; Touvron et al., 2023) series, and adapt them for translation tasks.",
+        "bbox": [
+            507,
+            695,
+            884,
+            821
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "One line of LLM-based methods focuses on zero-shot or few-shot translation by incorporating incontext learning(Hendy et al., 2023). By conditioning the LLM on a source sentence, the model can generate translations in the target language without explicitly using parallel data. This approach",
+        "bbox": [
+            507,
+            825,
+            882,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "640",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/73f436e14355582f68c0076b18f2226258b42498f47b7b07c4ef7bfa9a5ff179.jpg",
+        "image_caption": [
+            "Figure 1: The overall of our approach. Stage 1: Secondary Pre-training using Extensive Monolingual Data. Stage 2: Continual Pre-training with Interlinear Text Format Documents. Stage 3: Leveraging Source-Language Consistent Instruction for Supervised Fine-Tuning. *It should be noted that Stage 1 is considered non-essential.*"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            139,
+            80,
+            374,
+            323
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/6a3dc8f701c6c3d629cf4e86011d2f56bc3eea3b6069fec21035ccb49d0e3865.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            383,
+            80,
+            615,
+            322
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/733b32e518371d86814b6b5a5aaa0d52670f85194e94c39dacc0e8f5f9cf944b.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            621,
+            80,
+            857,
+            322
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "has shown promising results in enabling translation for language pairs with limited or no parallel resources.",
+        "bbox": [
+            112,
+            401,
+            487,
+            447
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Another approach involves using a small amount of high-quality bilingual parallel data to construct translation-guiding instructions. These instructions explicitly define the translation behavior by providing source-language consistent cues during the supervised fine-tuning (SFT) process. By utilizing these specially crafted instructions, the LM can be fine-tuned to perform translation more accurately and robustly.",
+        "bbox": [
+            112,
+            450,
+            487,
+            593
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Overall, LLM-based methods present alternative approaches to machine translation, exploring the potential of leveraging pre-trained models and incorporating limited parallel resources or high-quality instructions to enhance translation quality.",
+        "bbox": [
+            112,
+            595,
+            487,
+            675
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 A New Training Recipe",
+        "text_level": 1,
+        "bbox": [
+            112,
+            687,
+            349,
+            703
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We propose an innovative training strategy to enhance the translation capabilities of Large Language Models. As shown in Figure 1, our approach consists of three stages: (1) Secondary Pre-training using Extensive Monolingual Data, (2) Continual Pre-training with Interlinear Text Format Documents, and (3) Leveraging Source-Language Consistent Instruction for Supervised Fine-Tuning.",
+        "bbox": [
+            112,
+            713,
+            487,
+            841
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Stage 1: Continual Pre-training using Extensive Monolingual Data",
+        "text_level": 1,
+        "bbox": [
+            112,
+            852,
+            455,
+            883
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this stage, our aim is to enhance the training of large language models (LLMs) by utilizing di",
+        "bbox": [
+            112,
+            890,
+            487,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "verse monolingual data. Currently, existing large models, such as Llama, are primarily pre-trained on English-centric corpora, resulting in relatively weaker comprehension and generation abilities in non-English languages. To expand the multilingual generation capabilities of LLMs, we suggest an incremental pre-training approach using extensive monolingual data.",
+        "bbox": [
+            507,
+            401,
+            882,
+            529
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "It is important to note that this stage primarily focuses on enhancing the intrinsic multilingual capacity of LLMs. While it is inherently related to machine translation tasks, it is not essential. On the one hand, we can select an existing LLM that already demonstrates robust multilingual capabilities as the base model for further training. On the other hand, even LLMs with limited multilingual support can benefit from the subsequent stages outlined in our methodology.",
+        "bbox": [
+            507,
+            531,
+            882,
+            692
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2 Stage 2: Continual Pre-training with Sentence-aligned Parallel Data",
+        "text_level": 1,
+        "bbox": [
+            507,
+            705,
+            843,
+            736
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Interlinear Text Format Interlinear Text Format are a specific type of parallel text resource that consists of source sentences and their corresponding translations displayed in a aligned format. Each source sentence is accompanied by its translation, typically presented word-by-word or phrase-by-phrase, to facilitate a clear interlingual correspondence. We build the Sentence-aligned Parallel Data into this format. See Figure 2.",
+        "bbox": [
+            507,
+            743,
+            882,
+            887
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Utilizing Interlinear Text Format offers several advantages for language understanding and trans",
+        "bbox": [
+            507,
+            889,
+            882,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "641",
+        "bbox": [
+            485,
+            928,
+            512,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/024c4842a7908aaf690c6e1f062fab15f2309385c15e626385f7be37fe5c018b.jpg",
+        "image_caption": [
+            "Figure 2: Interlinear Text Format Documents"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            121,
+            90,
+            477,
+            311
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "lation tasks. Firstly, these data provide explicit linguistic alignment at a fine-grained level, enabling the model to capture syntactic and semantic correspondences across languages. This aligns closely with the goals of machine translation, as it facilitates accurate encoding of source language information and improves the quality of generated translations. Additionally, interlinear data contributes to the learning of interlingual representations, allowing the model to better understand the relationship and transferability between languages.",
+        "bbox": [
+            112,
+            366,
+            489,
+            542
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Continual Pre-training To leverage the benefits of Interlinear Text Documents, we propose a Continual Pre-training strategy based on the LoRA(Hu et al., 2021) (Low-Rank Adaptation of Large Language Models) framework. LoRA is a robust and effective pre-training approach for language models, introduced in recent research.",
+        "bbox": [
+            112,
+            552,
+            489,
+            663
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "By leveraging the inherent alignment information present in Interlinear Text Documents, the model learns to align and generate translations that maintain syntactic and semantic consistency with the source sentences. This continual training process allows the model to progressively improve its ability to capture cross-lingual correspondences, resulting in enhanced translation quality.",
+        "bbox": [
+            112,
+            664,
+            489,
+            793
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3 Stage 3: Leveraging Source-Language Consistent Instruction for Supervised Fine-Tuning",
+        "text_level": 1,
+        "bbox": [
+            112,
+            804,
+            460,
+            852
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Source-Language Consistent Instruction In the field of machine translation, \"Source-Language Consistent Instruction\" refers to the practice of constructing translation instructions that maintain",
+        "bbox": [
+            112,
+            857,
+            489,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "consistency with the source language, aiming to achieve better results. This approach involves generating instructions that are closely related to the source language. By providing more accurate and clear guidance for supervised fine-tuning of models, this technique enhances translation quality.",
+        "bbox": [
+            507,
+            84,
+            884,
+            181
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/dd82a0a7d282247a7d349f9fd5efa33054e6d5b87d850d43ffe79ca6a063b03c.jpg",
+        "image_caption": [
+            "(b) Source-Language Consistent Instruction",
+            "Figure 3: Instruction Format"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            552,
+            193,
+            842,
+            326
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To illustrate this concept, let's consider translations in English  $\\Leftrightarrow$  Chinese and English  $\\Leftrightarrow$  German. Traditional approaches typically employ a standardized English-Fixed instruction such as \"Translate this sentence from the source language to the target language:\". However, in Source-Language Consistent Instruction, the instruction varies based on the language pair. For English-to-Chinese translation, the instruction would be \"把这句话从中文翻译成英文：\" (Translate this sentence from Chinese to English). Similarly, for German-to-English translation, the instruction would be \"Übersetzen Sie die folgenden Sätze vom Deutschen ins Englische:\" (Translate the following sentences from German to English). By utilizing language-specific instructions, there is a semantic consistency established between the instruction and the source language, resulting in clearer and more accurate guidance. As shown in Figure 3.",
+        "bbox": [
+            507,
+            394,
+            884,
+            700
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Supervised Instruction Fine-Tuning Supervised Instruction Fine-Tuning for machine translation tasks incorporates two pivotal aspects. Firstly, akin to the earlier phase of Continual Pre-training, we employ LoRA(Hu et al., 2021) to finely tune specific parameters of Language Learning Models (LLMs), thereby enhancing their efficiency. LoRA(Hu et al., 2021) plays a crucial role in preventing model overfitting and leads to notable performance improvements. With this approach, we judiciously fine-tune a subset of model parameters using low-rank updates, striking a delicate balance between model adaptation and computational effi",
+        "bbox": [
+            507,
+            712,
+            884,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "642",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "ciency.",
+        "bbox": [
+            112,
+            85,
+            171,
+            99
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Secondly, as emphasized in prior studies(Zhou et al., 2023; Maillard et al., 2023; Xu et al., 2023), LLMs exhibit benefits from a limited yet high-quality dataset. To ensure optimal data quality during the fine-tuning process, we leverage exceptional data sources. In line with previous research, we make use of meticulously curated human-written datasets derived from the WMT test data. These datasets undergo rigorous quality control measures, rendering them an ideal choice for fine-tuning purposes.",
+        "bbox": [
+            112,
+            105,
+            489,
+            282
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4 Experiments",
+        "text_level": 1,
+        "bbox": [
+            112,
+            309,
+            260,
+            326
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1 Datasets and Evaluation Metrics",
+        "text_level": 1,
+        "bbox": [
+            112,
+            346,
+            416,
+            361
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The overall data statistics are shown in Table 5 of Appendix A. For Stage 2, we utilized the WMT bilingual training dataset consisting of English  $\\Leftrightarrow$  German (En  $\\Leftrightarrow$  De) and English  $\\Leftrightarrow$  Chinese (En  $\\Leftrightarrow$  Zh) sentence pairs. The En  $\\Leftrightarrow$  De dataset comprised approximately 4.5 million pairs, while the En  $\\Leftrightarrow$  Zh dataset contained around 25 million pairs. Due to the higher number of En  $\\Leftrightarrow$  Zh pairs compared to En  $\\Leftrightarrow$  De, we sampled 4.5 million En  $\\Leftrightarrow$  Zh pairs for our experiments. Overall, the combined dataset contained nearly 1 billion tokens.",
+        "bbox": [
+            112,
+            376,
+            489,
+            568
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "For Stage 3, we employed the newstest2017-2020 dataset for both  $\\mathrm{En} \\Leftrightarrow \\mathrm{Zh}$  and  $\\mathrm{En} \\Leftrightarrow \\mathrm{De}$  translation tasks. This dataset included a total of 37.6 thousand sentence pairs for each language direction. To ensure consistency across the source language and target language, we organize these sentence pairs into Source-Language Consistent Instructions.",
+        "bbox": [
+            112,
+            574,
+            489,
+            687
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We additionally incorporate the test sets from the WMT22 competition, which are carefully curated to include more recent content from diverse domains such as news, social media, e-commerce, and conversations. The test sets for the  $\\mathrm{De} \\Rightarrow \\mathrm{En}$ ,  $\\mathrm{En} \\Rightarrow \\mathrm{De}$ ,  $\\mathrm{Zh} \\Rightarrow \\mathrm{En}$ , and  $\\mathrm{En} \\Rightarrow \\mathrm{Zh}$  tasks consist of 1984, 2037, 1875, and 2037 samples, respectively.",
+        "bbox": [
+            112,
+            690,
+            489,
+            804
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "For automatic evaluation, we utilize Sacre-BLEU, which implements BLEU(Papineni et al., 2002), and COMET(Rei et al., 2020) from Unbabel/wmt22-comet-da. SacreBLEU calculates similarity based on n-gram matching, while COMET leverages cross-lingual pretrained models for evaluation.",
+        "bbox": [
+            112,
+            808,
+            490,
+            920
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2 Setup",
+        "text_level": 1,
+        "bbox": [
+            509,
+            84,
+            601,
+            99
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We conducted our experiments using HuggingFace Transformers with open-source LLMs from the LLaMA(Touvron et al., 2023) family. Specifically, we utilized LLaMA2-7b with matched parameters as our foundation model. Additionally, we included LLaMA2-13b to explore the impact of different model sizes.",
+        "bbox": [
+            507,
+            107,
+            884,
+            218
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Due to computational constraints, we did not reproduce the foundational experiments from Stage 1. After Stage 1, we selected Chinese-LLaMA2(Cui et al., 2023) as our new foundation model. Chinese-LLaMA2 is an extended and optimized version of Llama-2, specifically tailored for Chinese language understanding and instruction comprehension. It incorporates a larger Chinese vocabulary and undergoes incremental pretraining on a large-scale Chinese dataset, which further enhances its semantic understanding capabilities.",
+        "bbox": [
+            507,
+            221,
+            885,
+            398
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "For Stage 2, Continual Pre-training, and Stage 3, Supervised Fine-Tuning, we referred to the hyperparameters employed in the Chinese-LLaMA2 project. During Stage 2, we trained the model for 1 epoch, and for Stage 3, we extended the training to 3 epochs.",
+        "bbox": [
+            507,
+            399,
+            882,
+            495
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Our experiments were conducted on 8 Nvidia GPUs with 64GB of memory each, utilizing DeepSpeed(Rasley et al., 2020) ZeRO 2 for model parallelization.",
+        "bbox": [
+            507,
+            497,
+            882,
+            560
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.3 Baselines",
+        "text_level": 1,
+        "bbox": [
+            507,
+            575,
+            628,
+            589
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We evaluate our method against two baseline categories, consistent with previous studies. Firstly, we compare our approach to prior studies that share our goal of leveraging LLMs for translation. Secondly, we benchmark against the current state-of-the-art (SoTA) translation models. It's important to note that this comparison may not be entirely fair due to disparities in training data and model architectures. For example, there is a significant contrast between the 175B GPT-3.5 model and our 7B model. Nevertheless, by utilizing the same test set, we can gain insights into our model's current performance.",
+        "bbox": [
+            507,
+            598,
+            882,
+            806
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In the category of prior similar work, we compare our model to the following approaches: BigTranslate(Yang et al., 2023), which extends LLaMA-1-13B to cover over 100 translation directions; TIM(Zeng et al., 2023), which leverages correct and incorrect examples to aid LLMs in learning translation; ParroT(Jiao et al., 2023), through",
+        "bbox": [
+            507,
+            809,
+            884,
+            921
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "643",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/641a81bb40c99a396ac96365d96835c2f68a4c499da81ff3db3522f7d50397d1.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Models</td><td colspan=\"2\">De→En</td><td colspan=\"2\">En→De</td><td colspan=\"2\">Zh→En</td><td colspan=\"2\">En→Zh</td></tr><tr><td>BLEU</td><td>COMET</td><td>BLEU</td><td>COMET</td><td>BLEU</td><td>COMET</td><td>BLEU</td><td>COMET</td></tr><tr><td colspan=\"9\">SoTA models</td></tr><tr><td>NLLB-54B(Team et al., 2022)</td><td>26.89</td><td>78.94</td><td>34.50</td><td>86.45</td><td>16.56</td><td>70.70</td><td>27.38</td><td>78.91</td></tr><tr><td>NLLB-54B Fine-tune</td><td>27.34</td><td>79.86</td><td>35.07</td><td>86.95</td><td>17.26</td><td>71.35</td><td>27.89</td><td>80.13</td></tr><tr><td>GPT-3.5-D, zero-shot</td><td>30.90</td><td>84.79</td><td>31.80</td><td>85.61</td><td>25.00</td><td>81.60</td><td>38.30</td><td>85.76</td></tr><tr><td>GPT-3.5-T, zero-shot</td><td>33.10</td><td>85.50</td><td>34.40</td><td>87.00</td><td>26.60</td><td>82.90</td><td>44.90</td><td>87.00</td></tr><tr><td>GPT-4</td><td>33.87</td><td>85.62</td><td>35.38</td><td>87.44</td><td>27.20</td><td>82.79</td><td>43.98</td><td>87.49</td></tr><tr><td colspan=\"9\">Prior Similar Studies</td></tr><tr><td>TIM-7B(Zeng et al., 2023)</td><td>27.91</td><td>82.80</td><td>25.59</td><td>82.56</td><td>19.33</td><td>75.46</td><td>19.33</td><td>75.46</td></tr><tr><td>Parrot-7B(Jiao et al., 2023)</td><td>29.80</td><td>83.00</td><td>26.10</td><td>81.60</td><td>20.20</td><td>75.90</td><td>30.30</td><td>80.30</td></tr><tr><td>SWIE-7B(Chen et al., 2023)</td><td>30.48</td><td>82.97</td><td>27.21</td><td>82.36</td><td>21.30</td><td>76.48</td><td>31.24</td><td>80.63</td></tr><tr><td>ALMA-7B(Xu et al., 2023)</td><td>29.56</td><td>83.95</td><td>30.31</td><td>85.59</td><td>23.64</td><td>79.78</td><td>36.48</td><td>85.05</td></tr><tr><td>Parrot-13B(Jiao et al., 2023)</td><td>31.10</td><td>83.60</td><td>28.10</td><td>82.60</td><td>21.70</td><td>76.70</td><td>31.70</td><td>81.00</td></tr><tr><td>BigTranslate-13B(Yang et al., 2023)</td><td>23.35</td><td>80.68</td><td>21.48</td><td>78.81</td><td>14.16</td><td>74.26</td><td>28.56</td><td>81.31</td></tr><tr><td>Bayling-13B(Zhang et al., 2023)</td><td>27.34</td><td>83.02</td><td>25.62</td><td>82.69</td><td>20.12</td><td>77.72</td><td>37.92</td><td>84.62</td></tr><tr><td>ALMA-13B(Xu et al., 2023)</td><td>31.14</td><td>84.56</td><td>31.47</td><td>85.62</td><td>25.46</td><td>80.21</td><td>39.84</td><td>85.96</td></tr><tr><td>Ours</td><td colspan=\"8\">Our Recipe with Backbone Model: LLaMA2(Touvron et al., 2023)</td></tr><tr><td>7B Stage3</td><td>30.02</td><td>84.09</td><td>25.40</td><td>82.30</td><td>20.59</td><td>76.18</td><td>30.60</td><td>80.40</td></tr><tr><td>7B Stage1,3*</td><td>25.20</td><td>78.32</td><td>12.50</td><td>69.19</td><td>20.90</td><td>76.40</td><td>35.00</td><td>84.32</td></tr><tr><td>7B Stage2,3</td><td>31.14</td><td>84.70</td><td>30.50</td><td>85.62</td><td>21.97</td><td>78.45</td><td>39.00</td><td>85.79</td></tr><tr><td>7B Stage1,2,3*</td><td>30.10</td><td>83.96</td><td>29.90</td><td>83.86</td><td>22.20</td><td>79.88</td><td>41.10</td><td>86.37</td></tr><tr><td>13B Stage3</td><td>31.70</td><td>84.39</td><td>28.80</td><td>83.87</td><td>21.40</td><td>77.68</td><td>35.90</td><td>84.23</td></tr><tr><td>13B Stage1,3*</td><td>26.13</td><td>78.65</td><td>12.79</td><td>72.23</td><td>21.40</td><td>78.28</td><td>37.34</td><td>85.27</td></tr><tr><td>13B Stage2,3</td><td>32.24</td><td>85.17</td><td>32.53</td><td>86.14</td><td>22.57</td><td>79.05</td><td>40.40</td><td>85.98</td></tr><tr><td>13B Stage1,2,3*</td><td>30.21</td><td>84.26</td><td>30.41</td><td>84.72</td><td>23.10</td><td>80.53</td><td>42.30</td><td>86.65</td></tr></table>",
+        "bbox": [
+            126,
+            80,
+            870,
+            432
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 1: The overall results. Note: * Due to computational constraints, we did not reproduce the foundational experiments from Stage 1, but instead directly utilized the Chinese-Llama2(Cui et al., 2023) that had undergone similar training. Since Chinese-Llama2(Cui et al., 2023) was only trained in Chinese during Stage 1, our main analysis about Stage 1 focuses on its performance in  $\\mathrm{{Zh}} \\Rightarrow  \\mathrm{{En}}$  and  $\\mathrm{{En}} \\Rightarrow  \\mathrm{{Zh}}$  translations.",
+        "bbox": [
+            112,
+            442,
+            882,
+            500
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "three types of instructions including translation instruction, contrastive instruction, and error-guided instruction, improves the translation performance of LLM after SFT; SWIE(Chen et al., 2023), which enhances LLMs in translation through instruction augmentation; BayLing(Zhang et al., 2023), which incorporates interactive translation instructions; and ALMA(Xu et al., 2023), a two-stage finetuning method that initially fine-tunes on monolingual data and subsequently on a small set of high-quality parallel data.",
+        "bbox": [
+            112,
+            524,
+            489,
+            702
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In the SoTA models category, we consider the following: the NLLB-54B(Team et al., 2022) model, the largest and best translation model released in the NLLB family; the zero-shot performance of GPT3.5-text-davinci-003 (GPT-3.5-D) and GPT-3.5-turbo-0301 (GPT-3.5-T). Additionally, we present the zero-shot results for GPT-4. For a fair comparison, we also compared the results of fine-tuning NLLB-54B model with 37.6k data in Stage 3. To evaluate these baselines, we employ the same test data and evaluation metrics, reporting BLEU(Papineni et al., 2002) and COMET(Rei et al., 2020) scores as provided in their respective",
+        "bbox": [
+            112,
+            712,
+            489,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "papers.",
+        "bbox": [
+            509,
+            526,
+            566,
+            539
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5 Results and Analysis",
+        "text_level": 1,
+        "bbox": [
+            507,
+            554,
+            722,
+            570
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "As shown in Table 1, overall, our results outperform all previous studies, NLLB-54B(Team et al., 2022), and GPT-3.5-D, except for a slight lag in  $\\mathrm{Zh}\\Rightarrow \\mathrm{En}$ . Even our 7B model surpasses the results of other works. Particularly in the  $\\mathrm{En}\\Rightarrow \\mathrm{Zh}$  direction, our BLEU score is approximately 2.5 higher than the previous state-of-the-art. These findings are a testament to the effectiveness of our approach.",
+        "bbox": [
+            507,
+            581,
+            884,
+            726
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.1 Assessing the Impact of Stage 1",
+        "text_level": 1,
+        "bbox": [
+            507,
+            738,
+            801,
+            753
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Just as mentioned earlier, we didn't specifically train Llama2 in Stage 1, but instead directly utilized the Chinese-Llama2(Cui et al., 2023) that had undergone similar training. Since Chinese-Llama2(Cui et al., 2023) was only trained in Chinese during Stage 1, our main analysis focuses on its performance in  $\\mathrm{Zh}\\Rightarrow \\mathrm{En}$  and  $\\mathrm{En}\\Rightarrow \\mathrm{Zh}$  translations.",
+        "bbox": [
+            507,
+            759,
+            884,
+            885
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "As shown in Table 1, our findings align with previous research conclusions that incremental train",
+        "bbox": [
+            507,
+            889,
+            884,
+            920
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "644",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "ing on monolingual data is beneficial. Furthermore, we discovered that this benefit primarily affects the target language in translation tasks. For example, we observed a significant improvement in the performance of the 7B model on the  $\\mathrm{En} \\Rightarrow \\mathrm{Zh}$  test set, where the BLEU score increased from 30.60 to 35.00, a substantial improvement of 4.4 points. However, the improvement in the  $\\mathrm{Zh} \\Rightarrow \\mathrm{En}$  direction was limited, indicating that the role of Stage 1 is to enhance generation rather than comprehension.",
+        "bbox": [
+            112,
+            84,
+            489,
+            244
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Additionally, we found that performing incremental training on only one monolingual dataset had disastrous effects on translation tasks in other languages. For example, on the  $\\mathrm{En} \\Rightarrow \\mathrm{De}$  test set, the BLEU score plummeted from 25.40 to 12.50. Therefore, for multilingual translation, it is crucial to conduct Stage 1 training on multiple languages.",
+        "bbox": [
+            112,
+            244,
+            489,
+            357
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.2 Measuring the Effectiveness of Stage 2",
+        "text_level": 1,
+        "bbox": [
+            112,
+            369,
+            463,
+            384
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "As shown in Table 1, Llama2(Touvron et al., 2023) demonstrates improved quality across various test sets after Stage 2 training. An interesting observation, considering Llama2 as a large model primarily focused on English, is that the enhancement in English-Other translations is particularly noteworthy after Stage 2 Training. For instance, the 7B model exhibits remarkable improvements in  $\\mathrm{En}\\Rightarrow \\mathrm{De}$ , with the BLEU score increasing from 25.40 to 30.50, and in  $\\mathrm{En}\\Rightarrow \\mathrm{Zh}$ , where it rises from 30.60 to 39.00. The magnitude of these improvements is quite significant. Encouragingly, there are also improvements observed in translations from other languages to English.",
+        "bbox": [
+            112,
+            388,
+            489,
+            615
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "An even more intriguing finding is that, as mentioned before, since Chinese-Llama2(Cui et al., 2023) only underwent Stage 1 training on Chinese, its translation performance substantially deteriorates in the  $\\mathrm{En} \\Rightarrow \\mathrm{De}$  direction. However, with the magical touch of Stage 2 training, these capabilities are miraculously restored. The 7B model, on  $\\mathrm{En} \\Rightarrow \\mathrm{De}$ , rebounds from 12.50 to 29.90, approaching the results of the original Llama2(Touvron et al., 2023). These outcomes effectively affirm the effectiveness of Stage 2.",
+        "bbox": [
+            112,
+            615,
+            489,
+            791
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "After considering the overall process, we are interested in understanding the impact of Stage 2 only. As mentioned before, LLMs typically include two main types of models: Foundation Models and Chat Models. Evaluating Stage 2 essentially assesses the Foundation Model by using an n-shot evaluation, which includes both zero-shot and 5-shot evaluations. We have noticed that zero-shot",
+        "bbox": [
+            112,
+            793,
+            489,
+            920
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "evaluations can occur hallucinations. Hence, we are presenting the results of the 5-shot evaluation in Table 2.",
+        "bbox": [
+            507,
+            84,
+            884,
+            131
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/2d70c67d07fcd42ccb8f212de3c91dcf8a115808e3fa376c62b88faf45d0c183.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Models</td><td colspan=\"2\">Zh→En</td><td colspan=\"2\">En→Zh</td></tr><tr><td></td><td>BLEU</td><td>COMET</td><td>BLEU</td><td>COMET</td></tr><tr><td>Baseline</td><td>20.63</td><td>76.32</td><td>29.96</td><td>79.34</td></tr><tr><td>+ Stage2</td><td>21.64</td><td>78.07</td><td>38.62</td><td>85.30</td></tr></table>",
+        "bbox": [
+            524,
+            141,
+            873,
+            202
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.3 Analyzing the Outcomes of Stage 3",
+        "text_level": 1,
+        "bbox": [
+            507,
+            273,
+            830,
+            288
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "To evaluate the effectiveness of Source-Language Consistent Instruction in Stage 3, we conducted a comparative experiment using English-Fixed Instruction. The results of the experiment are presented in Table 3. It is evident that in the  $\\mathrm{En}\\Rightarrow \\mathrm{De}$  and  $\\mathrm{De}\\Rightarrow \\mathrm{En}$  directions, the performance of these two types of instructions is quite similar. However, in the  $\\mathrm{Zh}\\Rightarrow \\mathrm{En}$  and  $\\mathrm{En}\\Rightarrow \\mathrm{Zh}$  directions, the use of Source-Language Consistent Instruction clearly outperforms.",
+        "bbox": [
+            505,
+            294,
+            884,
+            455
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/67c454576b4166eea6a3a0945b9c197d954f7ca1ad7ec9cbfd2549425fdb2919.jpg",
+        "table_caption": [
+            "Table 2: Results of the five-shot results based on Llama2-7B(Touvron et al., 2023) model."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Models</td><td>De⇒En</td><td>En⇒De</td><td>Zh⇒En</td><td>En⇒Zh</td></tr><tr><td>Stage 3</td><td>30.02</td><td>25.40</td><td>20.59</td><td>30.60</td></tr><tr><td>w/o</td><td>30.40</td><td>25.20</td><td>18.39</td><td>28.30</td></tr><tr><td>Stage2,3</td><td>31.14</td><td>30.50</td><td>21.91</td><td>39.00</td></tr><tr><td>w/o</td><td>31.00</td><td>30.23</td><td>18.93</td><td>38.69</td></tr></table>",
+        "bbox": [
+            517,
+            466,
+            878,
+            544
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Results of the comparative experiments based on Llama2-7B(Touvron et al., 2023) model. [w/o] means using English-Fixed Instruction.",
+        "bbox": [
+            507,
+            552,
+            882,
+            596
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We believe that the similarity between English and German, as they belong to the same language family, contributes to the lack of noticeable differences. However, when dealing with cross-language pairs, employing Source-Language Consistent Instruction further enhances the translation quality.",
+        "bbox": [
+            507,
+            613,
+            884,
+            709
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.4 Comparing the Difference with ALMA",
+        "text_level": 1,
+        "bbox": [
+            507,
+            722,
+            860,
+            738
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We have noticed that our work shares some similarities with ALMA(Xu et al., 2023) in terms of the process, involving Continual Pre-training followed by Supervised Fine-Tuning. However, there are notable differences between our approaches.",
+        "bbox": [
+            505,
+            744,
+            882,
+            824
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "ALMA suggests that the impact of bilingual data is reduced in the era of large models. In contrast, we firmly believe in the continued strength of bilingual data and its application in Continual Pre-training through Interlinear Text Format Documents. While ALMA acknowledges the effective",
+        "bbox": [
+            507,
+            825,
+            884,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "645",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "ness of conducting Continual Pre-training on monolingual data, we have also validated this finding in our own work and reached the same conclusion. However, it is important to note that our approach primarily enhances the multilingual generation capability of the large model itself, rather than being specifically tailored to translation tasks. Furthermore, ALMA utilizes a significantly larger training dataset, ranging from 13B to 20B, compared to our own.",
+        "bbox": [
+            112,
+            84,
+            489,
+            244
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6 Ablation Study: What if we directly employ a large quantity of translation data for SFT?",
+        "text_level": 1,
+        "bbox": [
+            112,
+            266,
+            473,
+            317
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Both Continual Pre-training and Supervised Fine-Tuning involve incremental training on the original model. However, if we skip Stage 2 training and directly utilize the translation data from Stage 2 as instruction data for SFT, i.e., conducting SFT directly with a substantial amount of translation data, will it yield consistent improvement?",
+        "bbox": [
+            112,
+            331,
+            489,
+            444
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/8918374d11aaa79a68a936712827753ddca6e9ec0fb60685517c0b490da91898.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Data Size</td><td>De→En</td><td>En→De</td><td>Zh→En</td><td>En→Zh</td></tr><tr><td>37.6K</td><td>30.02</td><td>25.40</td><td>20.59</td><td>30.60</td></tr><tr><td>400K</td><td>30.20</td><td>25.60</td><td>18.49</td><td>31.74</td></tr><tr><td>4,000K</td><td>30.66</td><td>25.12</td><td>20.77</td><td>32.22</td></tr></table>",
+        "bbox": [
+            122,
+            458,
+            484,
+            520
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 4: Results of the ablation experiments based on Llama2-7B(Touvron et al., 2023) model under different Stage 3 data size.",
+        "bbox": [
+            112,
+            529,
+            487,
+            574
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "To address this question, we conducted an ablation experiment. Our Stage 3 training data consisted of  $37.6\\mathrm{k}$  samples. Randomly selecting and merging some data from the Stage 2 training data with the Stage 3 training data, we created three sets:  $37.6\\mathrm{K},400\\mathrm{K}$ , and  $4,000\\mathrm{K}$ . The experimental results are presented in Table 4.",
+        "bbox": [
+            112,
+            594,
+            489,
+            707
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We found that augmenting the training data in Stage 3 slightly improved translation quality for certain test sets. This indicates that a small amount of high-quality data is sufficient for the SFT stage.",
+        "bbox": [
+            112,
+            709,
+            489,
+            774
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Now, our focus is solely on the translation task. However, if we were conducting multi-task SFT, it is unlikely that other tasks would have as extensive data as machine translation. Therefore, using a large amount of translation data during SFT would result in the problem of imbalanced data distribution with other tasks. Hence, the optimal approach would still be to utilize this substantial data during the earlier stage of Continual Pre-training.",
+        "bbox": [
+            112,
+            776,
+            490,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "7 Conclusions",
+        "text_level": 1,
+        "bbox": [
+            509,
+            83,
+            650,
+            99
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this study, we have introduced a novel paradigm for enhancing the translation capabilities of large language models in machine translation tasks. Our three-stage approach, including Secondary Pretraining using Extensive Monolingual Data, Continual Pre-training with Interlinear Text Format Documents, and Leveraging Source-Language Consistent Instruction for Supervised Fine-Tuning, addresses the limitations of previous strategies and offers notable improvements in translation quality. We emphasize the significance of pre-training stages in enhancing LLMs' cross-lingual alignment abilities and the effectiveness of using a smaller but high-quality set of bilingual data during supervised fine-tuning. Notably, Stage2, which involves Continual Pre-training with Interlinear Text Format Documents, stands out as a highly efficient method, requiring minimal training data. Furthermore, aligning the instructional setting with the source language during supervised fine-tuning, as observed in Stage3, yields positive effects. The findings from this paper contribute to advancing the field of machine translation and offer valuable insights for optimizing the translation capabilities of large language models. Future research can explore additional language pairs, alternative data augmentation techniques, and different pre-training strategies to further refine our proposed paradigm.",
+        "bbox": [
+            505,
+            110,
+            885,
+            561
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "8 Limitations",
+        "text_level": 1,
+        "bbox": [
+            507,
+            573,
+            645,
+            589
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Despite notable contributions, this study has certain limitations. Firstly, the proposed method exhibits slightly reduced performance in the  $\\mathrm{Zh} \\Rightarrow \\mathrm{En}$  translation direction, necessitating further analysis and improvements. Secondly, the presence of illusionary translations within large models was observed but not extensively explored. Future research should delve deeper into this phenomenon. Lastly, while the paper primarily focuses on SFT for machine translation, opportunities exist to explore SFT techniques in diverse contexts such as style translation and colloquial translation. Addressing these limitations would further enhance the effectiveness and applicability of the proposed methods.",
+        "bbox": [
+            507,
+            600,
+            885,
+            841
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            509,
+            869,
+            610,
+            885
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind",
+        "bbox": [
+            509,
+            892,
+            882,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "646",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. CoRR, abs/2005.14165.",
+            "Yijie Chen, Yijin Liu, Fandong Meng, Yufeng Chen, Jinan Xu, and Jie Zhou. 2023. Improving translation faithfulness of large language models via augmenting instructions.",
+            "Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. 2023. Palm: Scaling language modeling with pathways. J. Mach. Learn. Res., 24:240:1-240:113.",
+            "Yiming Cui, Ziqing Yang, and Xin Yao. 2023. Efficient and effective text encoding for chinese llama and alpaca. arXiv preprint arXiv:2304.08177.",
+            "Marjan Ghazvininejad, Omer Levy, Yinhan Liu, and Luke Zettlemoyer. 2019. Mask-predict: Parallel decoding of conditional masked language models. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, November 3-7, 2019, pages 6111-6120. Association for Computational Linguistics.",
+            "Jiatao Gu, James Bradbury, Caiming Xiong, Victor O. K. Li, and Richard Socher. 2018. Non-autoregressive neural machine translation. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net.",
+            "Jiaxin Guo, Minghan Wang, Daimeng Wei, Hengchao Shang, Yuxia Wang, Zongyao Li, Zhengzhe Yu, Zhanglin Wu, Yimeng Chen, Chang Su, Min Zhang,"
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            920
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Lizhi Lei, Shimin Tao, and Hao Yang. 2021. Self-distillation mixup training for non-autoregressive neural machine translation. CoRR, abs/2112.11640.",
+            "Amr Hendy, Mohamed Abdelrehim, Amr Sharaf, Vikas Raunak, Mohamed Gabr, Hitokazu Matsushita, Young Jin Kim, Mohamed Affify, and Hany Hassan Awadalla. 2023. How good are GPT models at machine translation? A comprehensive evaluation. CoRR, abs/2302.09210.",
+            "Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models.",
+            "Wenxiang Jiao, Jen tse Huang, Wenxuan Wang, Zhiwei He, Tian Liang, Xing Wang, Shuming Shi, and Zhaopeng Tu. 2023. Parrot: Translating during chat using large language models tuned with human translation and feedback.",
+            "Jean Maillard, Cynthia Gao, Elahe Kalbassi, Kaushik Ram Sadagopan, Vedanuj Goswami, Philipp Koehn, Angela Fan, and Francisco Guzmán. 2023. Small data, big impact: Leveraging minimal data for effective machine translation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2023, Toronto, Canada, July 9-14, 2023, pages 2740-2756. Association for Computational Linguistics.",
+            "Kishore Papineni, Salim Roukos, Todd Ward, and Wei Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, July 6-12, 2002, Philadelphia, PA, USA, pages 311-318. ACL.",
+            "Jeff Rasley, Samyam Rajbhandari, Olatunjri Ruwase, and Yuxiong He. 2020. Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In KDD '20: The 26th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Virtual Event, CA, USA, August 23-27, 2020, pages 3505-3506. ACM.",
+            "Ricardo Rei, Craig Stewart, Ana C. Farinha, and Alon Lavie. 2020. COMET: A neural framework for MT evaluation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020, pages 2685-2702. Association for Computational Linguistics.",
+            "Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Improving neural machine translation models with monolingual data.",
+            "NLLB Team, Marta R. Costa-jussa, James Cross, Onur Celebi, Maha Elbayad, Kenneth Heafield, Kevin Hefernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, Anna Sun, Skyler Wang, Guillaume Wenzek, Al Youngblood, Bapi Akula, Loic Barrault, Gabriel Mejia Gonzalez, Prangthip Hansanti,"
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            920
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "647",
+        "bbox": [
+            485,
+            928,
+            515,
+            939
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "John Hoffman, Semarley Jarrett, Kaushik Ram Sadagopan, Dirk Rowe, Shannon Spruit, Chau Tran, Pierre Andrews, Necip Fazil Ayan, Shruti Bhosale, Sergey Edunov, Angela Fan, Cynthia Gao, Vedanuj Goswami, Francisco Guzmán, Philipp Koehn, Alexandre Mourachko, Christophe Ropers, Safiyyah Saleem, Holger Schwenk, and Jeff Wang. 2022. No language left behind: Scaling human-centered machine translation.",
+        "bbox": [
+            131,
+            85,
+            489,
+            203
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. 2023. Llama 2: Open foundation and finetuned chat models.",
+        "bbox": [
+            115,
+            212,
+            489,
+            513
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5998-6008.",
+        "bbox": [
+            115,
+            523,
+            489,
+            617
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Minghan Wang, Jiaxin Guo, Yuxia Wang, Yimeng Chen, Chang Su, Daimeng Wei, Min Zhang, Shimin Tao, and Hao Yang. 2021. HI-CMLM: improve CMLM with hybrid decoder input. In Proceedings of the 14th International Conference on Natural Language Generation, INLG 2021, Aberdeen, Scotland, UK, 20-24 September, 2021, pages 167-171. Association for Computational Linguistics.",
+        "bbox": [
+            115,
+            626,
+            489,
+            731
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Daimeng Wei, Zhanglin Wu, Hengchao Shang, Zongyao Li, Minghan Wang, Jiaxin Guo, Xiaoyu Chen, Zhengzhe Yu, and Hao Yang. 2023. Text style transfer back-translation.",
+        "bbox": [
+            115,
+            741,
+            489,
+            793
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Haoran Xu, Young Jin Kim, Amr Sharaf, and Hany Hassan Awadalla. 2023. A paradigm shift in machine translation: Boosting translation performance of large language models.",
+        "bbox": [
+            115,
+            804,
+            489,
+            857
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Wen Yang, Chong Li, Jiajun Zhang, and Chengqing Zong. 2023. Bigtranslate: Augmenting large language models with multilingual translation capability over 100 languages.",
+        "bbox": [
+            115,
+            866,
+            489,
+            920
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Zhengzhe Yu, Jiaxin Guo, Minghan Wang, Daimeng Wei, Hengchao Shang, Zongyao Li, Zhanglin Wu, Yuxia Wang, Yimeng Chen, Chang Su, Min Zhang, Lizhi Lei, Shimin Tao, and Hao Yang. 2021. Jointtraining on symbiosis networks for deep nerval machine translation models. CoRR, abs/2112.11642.",
+        "bbox": [
+            509,
+            85,
+            884,
+            164
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Jiali Zeng, Fandong Meng, Yongjing Yin, and Jie Zhou. 2023. Tim: Teaching large language models to translate with comparison.",
+        "bbox": [
+            509,
+            174,
+            884,
+            214
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Shaolei Zhang, Qingkai Fang, Zhuocheng Zhang, Zhengrui Ma, Yan Zhou, Langlin Huang, Mengyu Bu, Shangtong Gui, Yunji Chen, Xilin Chen, and Yang Feng. 2023. Bayling: Bridging cross-lingual alignment and instruction following through interactive translation for large language models.",
+        "bbox": [
+            509,
+            223,
+            884,
+            303
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Chunting Zhou, Pengfei Liu, Puxin Xu, Srini Iyer, Jiao Sun, Yuning Mao, Xuezhe Ma, Avia Efrat, Ping Yu, Lili Yu, Susan Zhang, Gargi Ghosh, Mike Lewis, Luke Zettlemoyer, and Omer Levy. 2023. LIMA: less is more for alignment. CoRR, abs/2305.11206.",
+        "bbox": [
+            509,
+            312,
+            884,
+            379
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "A Appendix A: Data Statistics",
+        "text_level": 1,
+        "bbox": [
+            509,
+            403,
+            786,
+            419
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Table 5 displays the comprehensive data statistics. For Stage 2, we utilized the WMT bilingual training dataset that includes English  $\\Leftrightarrow$  German (En  $\\Leftrightarrow$  De) and English  $\\Leftrightarrow$  Chinese (En  $\\Leftrightarrow$  Zh) sentence pairs. The En  $\\Leftrightarrow$  De dataset comprised approximately 4.5 million pairs, while for the En  $\\Leftrightarrow$  Zh dataset, we randomly sampled an equivalent number of pairs from the total 25 million pairs. In total, the combined dataset contained close to 1B tokens.",
+        "bbox": [
+            507,
+            429,
+            884,
+            571
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Moving on to Stage 3, we utilized the newstest2017-2020 dataset for both the  $\\mathrm{En} \\Leftrightarrow \\mathrm{De}$  and  $\\mathrm{En} \\Leftrightarrow \\mathrm{Zh}$  translation tasks. This dataset comprised 37.6 thousand sentence pairs for each language direction. To maintain coherence between the source and target languages, we categorized these sentence pairs into Source-Language Consistent Instructions.",
+        "bbox": [
+            507,
+            573,
+            884,
+            700
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "B Appendix B: Stage 2 Training Data Sample",
+        "text_level": 1,
+        "bbox": [
+            509,
+            713,
+            848,
+            747
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Figure 4 displays samples of the Stage 2 training data.",
+        "bbox": [
+            507,
+            756,
+            880,
+            785
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "648",
+        "bbox": [
+            485,
+            928,
+            515,
+            939
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "table",
+        "img_path": "images/829dcd4b8e25c1bc2e6460d96b418d701c91734de13178a4c59892fa1d6afb3c.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Stage</td><td>Data</td><td>Description</td></tr><tr><td>Stage 1</td><td>120G text</td><td>120G Chinese text described in Chinese-LLaMA2(Cui et al., 2023)</td></tr><tr><td>Stage 2</td><td>1B tokens</td><td>1B tokens including four directions: En↔De and En↔Zh. Each direction includes 4.5 million pairs.</td></tr><tr><td>Stage 3</td><td>37.6k pairs</td><td>37.6k pairs combined wmt newstest2017-2020 testset with all four directions.</td></tr></table>",
+        "bbox": [
+            149,
+            186,
+            848,
+            265
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Table 5: The overall data statistics.",
+        "bbox": [
+            371,
+            274,
+            623,
+            288
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "<s> But buses are faster and more frequent. \\n Aber Buses sind Schneller und haufter. </s> <s> 因此, 公布的历史成本不反映非消耗性财产充分而准确的估价; \\n Accordingly, the historical costs disclosed did not reflect the full and accurate valuation of non-expendable property; </s> Wir haben heute dem Bericht Ferber zugestimmt, obwohl er eine Grundlage für den Kauf der Parlamentsgebäude in Straßburg bildet. Unsere Zustimmung stellt aber keine Vorfestlegung auf den Sitz des Europäischen Parlaments dar. \\n We voted in favour of the Ferber report today, even though it forms a basis for the purchase of the Parliament building in Strasbourg. However, our approval does not constitute a prior decision regarding the seat of the European Parliament. </s> <s> By introducing greater flexibility in how the Fund is used and by reducing the number of redundancies from 1000 to 500, it will become an ever more effective instrument for helping to tackle the effects of the economic down-turn. \\n Durch die Einführung von mehr Flexibilität in der Nutzung der Fonds und durch die Reduzierung der Zahl der Entlassungen von 1000 auf 500 wird er ein noch effizienteres Instrument zur Bekämpfung der Folgen des Wirtschaftssabschwungs werden. </s> <s> During the period from 1 January to 30 June 2010, contributions totalling $19,299.39 were received from Angola ($10,000) and Congo ($9,299.39). \\n 2010年1月1日至6月30日期间,从安哥拉(10000美元)和刚果(9299.39美元)共收到19299.39美元捐款。 </s> <s> ...",
+        "bbox": [
+            297,
+            514,
+            695,
+            778
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Figure 4: Stage 2 Training Data Sample",
+        "bbox": [
+            361,
+            796,
+            636,
+            812
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "649",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 10
+    }
+]

2024/A Novel Paradigm Boosting Translation Capabilities of Large Language Models/8eb804a5-0f10-4fdb-af4a-6a552f7a4c4e_model.json

@@ -0,0 +1,1894 @@
+[
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.127,
+                0.09,
+                0.872,
+                0.13
+            ],
+            "angle": 0,
+            "content": "A Novel Paradigm Boosting Translation Capabilities of Large Language Models"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.342,
+                0.148,
+                0.662,
+                0.165
+            ],
+            "angle": 0,
+            "content": "Jiaxin Guo*, Hao Yang†, Zongyao Li"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.304,
+                0.165,
+                0.7,
+                0.181
+            ],
+            "angle": 0,
+            "content": "Daimeng Wei, Hengchao Shang, Xiaoyu Chen"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.304,
+                0.182,
+                0.703,
+                0.198
+            ],
+            "angle": 0,
+            "content": "{jiaxinguo1,yanghao30,lizongyao} \\(@\\) huawei.com"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.263,
+                0.199,
+                0.743,
+                0.214
+            ],
+            "angle": 0,
+            "content": "{weidaimeng,shanghengchao,chenxiaoyu35}@huawei.com"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.293,
+                0.215,
+                0.711,
+                0.23
+            ],
+            "angle": 0,
+            "content": "Huawei Translation Services Center, Beijing, China"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.261,
+                0.261,
+                0.341,
+                0.276
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.144,
+                0.29,
+                0.461,
+                0.889
+            ],
+            "angle": 0,
+            "content": "This paper presents a study on strategies to enhance the translation capabilities of large language models (LLMs) in the context of machine translation (MT) tasks. The paper proposes a novel paradigm consisting of three stages: Secondary Pre-training using Extensive Monolingual Data, Continual Pre-training with Interlinear Text Format Documents, and Leveraging Source-Language Consistent Instruction for Supervised Fine-Tuning. Previous research on LLMs focused on various strategies for supervised fine-tuning (SFT), but their effectiveness has been limited. While traditional machine translation approaches rely on vast amounts of parallel bilingual data, our paradigm highlights the importance of using smaller sets of high-quality bilingual data. We argue that the focus should be on augmenting LLMs' cross-lingual alignment abilities during pre-training rather than solely relying on extensive bilingual data during SFT. Experimental results conducted using the Llama2(Touvron et al., 2023) model, particularly on Chinese-Llama2(Cui et al., 2023) after monolingual augmentation, demonstrate the improved translation capabilities of LLMs. A significant contribution of our approach lies in Stage2: Continual Pre-training with Interlinear Text Format Documents, which requires less than 1B training data, making our method highly efficient. Additionally, in Stage3, we observed that setting instructions consistent with the source language benefits the supervised fine-tuning process. Experimental results demonstrate that our approach surpasses previous work and achieves superior performance compared to models such as NLLB-54B(Team et al., 2022) and GPT3.5-text-davinci-003, despite having a significantly smaller parameter count of only 7B or 13B. This achievement establishes our method as a pioneering strategy in the field of machine translation."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.511,
+                0.261,
+                0.655,
+                0.276
+            ],
+            "angle": 0,
+            "content": "1 Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.29,
+                0.886,
+                0.628
+            ],
+            "angle": 0,
+            "content": "Translation capabilities of large language models (LLMs)(Brown et al., 2020; Chowdhery et al., 2023; Touvron et al., 2023) in machine translation (MT) tasks have been explored extensively in previous research(Jiao et al., 2023; Zeng et al., 2023; Chen et al., 2023; Xu et al., 2023; Yang et al., 2023; Zhang et al., 2023). However, achieving significant improvements in translation quality through supervised fine-tuning (SFT) strategies has proven challenging. Traditionally, machine translation relies on vast amounts of parallel bilingual data, but SFT only requires a small amount of high-quality bilingual data, highlighting a crucial distinction. It is a naive approach to consider using vast quantities of parallel bilingual data during SFT. However, experiments have shown that increasing the data volume yields limited improvements in quality and even leads to performance degradation on certain test sets. Thus, the question arises: are extensive parallel bilingual data useless in SFT, or are they being misused?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.63,
+                0.885,
+                0.774
+            ],
+            "angle": 0,
+            "content": "In this paper, we propose a novel training paradigm, consisting of three stages, to boost the translation capabilities of LLMs. Our contributions include refining the training strategy for downstream tasks and emphasizing the enhancement of LLMs' cross-lingual alignment abilities during pretraining. These contributions address the limitations observed in previous research. Our training paradigm comprises the following stages:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.776,
+                0.885,
+                0.904
+            ],
+            "angle": 0,
+            "content": "Stage 1: Continual Pre-training using Extensive Monolingual Data. Consistent with previous findings(Xu et al., 2023), we validate the effectiveness of monolingual data augmentation. Specifically, we perform SFT on a Chinese-Llama2(Cui et al., 2023) model, which undergoes monolingual data augmentation, thereby demonstrating improved translation capabilities."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.527,
+                0.906,
+                0.882,
+                0.921
+            ],
+            "angle": 0,
+            "content": "Stage 2: Continual Pre-training with"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.138,
+                0.895,
+                0.24,
+                0.906
+            ],
+            "angle": 0,
+            "content": "*Co-first Author"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.138,
+                0.907,
+                0.24,
+                0.919
+            ],
+            "angle": 0,
+            "content": "†Co-First author"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.138,
+                0.895,
+                0.24,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.486,
+                0.929,
+                0.516,
+                0.941
+            ],
+            "angle": 0,
+            "content": "639"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.228,
+                0.946,
+                0.771,
+                0.958
+            ],
+            "angle": 0,
+            "content": "Findings of the Association for Computational Linguistics: NAACL 2024, pages 639-649"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.293,
+                0.96,
+                0.706,
+                0.972
+            ],
+            "angle": 0,
+            "content": "June 16-21, 2024 ©2024 Association for Computational Linguistics"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.49,
+                0.246
+            ],
+            "angle": 0,
+            "content": "Sentence-aligned Parallel Data. We construct interlinear text format from sentence-aligned bilingual parallel data and utilize them for continual pre-training of LLMs. Experimental results demonstrate the critical importance of this stage, resulting in a significant improvement in translation quality, particularly for English-Other translations. Stage 2 stands as a pivotal contribution in our paper, requiring less than 1B training data, thereby enhancing training efficiency."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.247,
+                0.49,
+                0.358
+            ],
+            "angle": 0,
+            "content": "Stage 3: Leveraging Source-Language Consistent Instruction for Supervised Fine-Tuning. In SFT, we discover that using instruction aligned with the source language of the translation notably improves performance. Leveraging source-language consistent instructions during SFT yields significant enhancements."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.361,
+                0.49,
+                0.489
+            ],
+            "angle": 0,
+            "content": "In summary, we introduce a three-stage training paradigm, highlighting the effectiveness of secondary pre-training, continual pre-training with interlinear text format documents, and leveraging source-language consistent instruction for supervised fine-tuning. These contributions address the limitations observed in previous research and pave the way for improved translation quality."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.504,
+                0.27,
+                0.519
+            ],
+            "angle": 0,
+            "content": "2 Related Work"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.533,
+                0.353,
+                0.547
+            ],
+            "angle": 0,
+            "content": "2.1 Large Language Models"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.555,
+                0.49,
+                0.699
+            ],
+            "angle": 0,
+            "content": "Foundation Model Foundation Model, a product of pre-training, is a prominent type of Large Language Model. It has gained substantial recognition in recent years for its impressive capabilities in natural language processing tasks. The most prevalent architectural framework for such models is the Transformer, which employs a series of self-attention mechanisms to process input text efficiently."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.701,
+                0.49,
+                0.845
+            ],
+            "angle": 0,
+            "content": "Among the state-of-the-art Large Language Models, notable examples include GPT-3(Brown et al., 2020) and Llama2(Touvron et al., 2023). These models have been widely lauded for their exceptional proficiency in understanding and generating natural language text. They showcase the remarkable potential of Foundation Models, pushing the boundaries of language processing and setting new benchmarks in various applications."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.858,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Instruct/Chat Model Instruct/Chat Model, a variant of Large Language Models, is specifically developed through the process of Supervised Fine-Tuning (SFT). Unlike Foundation Models, which"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.885,
+                0.149
+            ],
+            "angle": 0,
+            "content": "are pre-trained, Instruct/Chat Models undergo additional supervised training to enhance their performance in specific tasks such as instruction following or conversational dialogue."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.151,
+                0.885,
+                0.262
+            ],
+            "angle": 0,
+            "content": "Supervised Fine-Tuning involves training the model on labeled datasets, where human annotators provide examples of desired input-output behavior. This approach enables Instruct/Chat Models to learn task-specific skills and exhibit improved performance in situations that require language understanding, generation, and interaction."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.264,
+                0.885,
+                0.36
+            ],
+            "angle": 0,
+            "content": "Noteworthy advancements have been observed in Instruct/Chat Models, with notable examples including models like ChatGPT. These models have exhibited remarkable outcomes in conversational scenarios, demonstrating their potential in enabling interactive and engaging human-like conversations."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.375,
+                0.762,
+                0.389
+            ],
+            "angle": 0,
+            "content": "2.2 Machine Translation Task"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.398,
+                0.884,
+                0.445
+            ],
+            "angle": 0,
+            "content": "Machine Translation Task refers to the process of automatically translating text from one language to another using computational methods."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.459,
+                0.885,
+                0.683
+            ],
+            "angle": 0,
+            "content": "Traditional Methods Traditional machine translation methods primarily rely on encoder-decoder(Vaswani et al., 2017) models, where an encoder converts the source language sentence and a decoder produces the translated sentence. These methods heavily depend on large bilingual parallel corpora for training, aligning source sentences with their corresponding translations. Data augmentation(Sennrich et al., 2016; Wei et al., 2023) is a common practice in traditional machine translation. Some studies(Gu et al., 2018; Ghazvininejad et al., 2019; Wang et al., 2021; Guo et al., 2021; Yu et al., 2021) also investigate transforming them into parallel architectures to speed up inference efficiency."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.696,
+                0.885,
+                0.822
+            ],
+            "angle": 0,
+            "content": "LLM-based Methods In recent years, Language Model (LM)-based approaches have gained attention in the field of machine translation. These approaches leverage pre-trained language models, such as the GPT (Generative Pre-trained Transformer)(Brown et al., 2020; Chowdhery et al., 2023; Touvron et al., 2023) series, and adapt them for translation tasks."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.826,
+                0.884,
+                0.922
+            ],
+            "angle": 0,
+            "content": "One line of LLM-based methods focuses on zero-shot or few-shot translation by incorporating incontext learning(Hendy et al., 2023). By conditioning the LLM on a source sentence, the model can generate translations in the target language without explicitly using parallel data. This approach"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.486,
+                0.929,
+                0.517,
+                0.941
+            ],
+            "angle": 0,
+            "content": "640"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.141,
+                0.081,
+                0.375,
+                0.324
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.384,
+                0.081,
+                0.616,
+                0.323
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.623,
+                0.082,
+                0.858,
+                0.323
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.334,
+                0.884,
+                0.378
+            ],
+            "angle": 0,
+            "content": "Figure 1: The overall of our approach. Stage 1: Secondary Pre-training using Extensive Monolingual Data. Stage 2: Continual Pre-training with Interlinear Text Format Documents. Stage 3: Leveraging Source-Language Consistent Instruction for Supervised Fine-Tuning. *It should be noted that Stage 1 is considered non-essential.*"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.403,
+                0.489,
+                0.448
+            ],
+            "angle": 0,
+            "content": "has shown promising results in enabling translation for language pairs with limited or no parallel resources."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.451,
+                0.489,
+                0.594
+            ],
+            "angle": 0,
+            "content": "Another approach involves using a small amount of high-quality bilingual parallel data to construct translation-guiding instructions. These instructions explicitly define the translation behavior by providing source-language consistent cues during the supervised fine-tuning (SFT) process. By utilizing these specially crafted instructions, the LM can be fine-tuned to perform translation more accurately and robustly."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.596,
+                0.489,
+                0.676
+            ],
+            "angle": 0,
+            "content": "Overall, LLM-based methods present alternative approaches to machine translation, exploring the potential of leveraging pre-trained models and incorporating limited parallel resources or high-quality instructions to enhance translation quality."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.688,
+                0.351,
+                0.705
+            ],
+            "angle": 0,
+            "content": "3 A New Training Recipe"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.714,
+                0.489,
+                0.842
+            ],
+            "angle": 0,
+            "content": "We propose an innovative training strategy to enhance the translation capabilities of Large Language Models. As shown in Figure 1, our approach consists of three stages: (1) Secondary Pre-training using Extensive Monolingual Data, (2) Continual Pre-training with Interlinear Text Format Documents, and (3) Leveraging Source-Language Consistent Instruction for Supervised Fine-Tuning."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.853,
+                0.456,
+                0.884
+            ],
+            "angle": 0,
+            "content": "3.1 Stage 1: Continual Pre-training using Extensive Monolingual Data"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.891,
+                0.489,
+                0.922
+            ],
+            "angle": 0,
+            "content": "In this stage, our aim is to enhance the training of large language models (LLMs) by utilizing di"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.403,
+                0.883,
+                0.53
+            ],
+            "angle": 0,
+            "content": "verse monolingual data. Currently, existing large models, such as Llama, are primarily pre-trained on English-centric corpora, resulting in relatively weaker comprehension and generation abilities in non-English languages. To expand the multilingual generation capabilities of LLMs, we suggest an incremental pre-training approach using extensive monolingual data."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.532,
+                0.884,
+                0.693
+            ],
+            "angle": 0,
+            "content": "It is important to note that this stage primarily focuses on enhancing the intrinsic multilingual capacity of LLMs. While it is inherently related to machine translation tasks, it is not essential. On the one hand, we can select an existing LLM that already demonstrates robust multilingual capabilities as the base model for further training. On the other hand, even LLMs with limited multilingual support can benefit from the subsequent stages outlined in our methodology."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.706,
+                0.844,
+                0.737
+            ],
+            "angle": 0,
+            "content": "3.2 Stage 2: Continual Pre-training with Sentence-aligned Parallel Data"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.744,
+                0.883,
+                0.888
+            ],
+            "angle": 0,
+            "content": "Interlinear Text Format Interlinear Text Format are a specific type of parallel text resource that consists of source sentences and their corresponding translations displayed in a aligned format. Each source sentence is accompanied by its translation, typically presented word-by-word or phrase-by-phrase, to facilitate a clear interlingual correspondence. We build the Sentence-aligned Parallel Data into this format. See Figure 2."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.89,
+                0.883,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Utilizing Interlinear Text Format offers several advantages for language understanding and trans"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.486,
+                0.929,
+                0.514,
+                0.941
+            ],
+            "angle": 0,
+            "content": "641"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.122,
+                0.091,
+                0.478,
+                0.312
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.147,
+                0.327,
+                0.456,
+                0.341
+            ],
+            "angle": 0,
+            "content": "Figure 2: Interlinear Text Format Documents"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.367,
+                0.49,
+                0.543
+            ],
+            "angle": 0,
+            "content": "lation tasks. Firstly, these data provide explicit linguistic alignment at a fine-grained level, enabling the model to capture syntactic and semantic correspondences across languages. This aligns closely with the goals of machine translation, as it facilitates accurate encoding of source language information and improves the quality of generated translations. Additionally, interlinear data contributes to the learning of interlingual representations, allowing the model to better understand the relationship and transferability between languages."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.553,
+                0.49,
+                0.664
+            ],
+            "angle": 0,
+            "content": "Continual Pre-training To leverage the benefits of Interlinear Text Documents, we propose a Continual Pre-training strategy based on the LoRA(Hu et al., 2021) (Low-Rank Adaptation of Large Language Models) framework. LoRA is a robust and effective pre-training approach for language models, introduced in recent research."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.665,
+                0.49,
+                0.794
+            ],
+            "angle": 0,
+            "content": "By leveraging the inherent alignment information present in Interlinear Text Documents, the model learns to align and generate translations that maintain syntactic and semantic consistency with the source sentences. This continual training process allows the model to progressively improve its ability to capture cross-lingual correspondences, resulting in enhanced translation quality."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.805,
+                0.462,
+                0.853
+            ],
+            "angle": 0,
+            "content": "3.3 Stage 3: Leveraging Source-Language Consistent Instruction for Supervised Fine-Tuning"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.858,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Source-Language Consistent Instruction In the field of machine translation, \"Source-Language Consistent Instruction\" refers to the practice of constructing translation instructions that maintain"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.885,
+                0.182
+            ],
+            "angle": 0,
+            "content": "consistency with the source language, aiming to achieve better results. This approach involves generating instructions that are closely related to the source language. By providing more accurate and clear guidance for supervised fine-tuning of models, this technique enhances translation quality."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.554,
+                0.195,
+                0.843,
+                0.327
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.625,
+                0.337,
+                0.789,
+                0.347
+            ],
+            "angle": 0,
+            "content": "(b) Source-Language Consistent Instruction"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.598,
+                0.362,
+                0.796,
+                0.376
+            ],
+            "angle": 0,
+            "content": "Figure 3: Instruction Format"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.395,
+                0.885,
+                0.701
+            ],
+            "angle": 0,
+            "content": "To illustrate this concept, let's consider translations in English \\(\\Leftrightarrow\\) Chinese and English \\(\\Leftrightarrow\\) German. Traditional approaches typically employ a standardized English-Fixed instruction such as \"Translate this sentence from the source language to the target language:\". However, in Source-Language Consistent Instruction, the instruction varies based on the language pair. For English-to-Chinese translation, the instruction would be \"把这句话从中文翻译成英文：\" (Translate this sentence from Chinese to English). Similarly, for German-to-English translation, the instruction would be \"Übersetzen Sie die folgenden Sätze vom Deutschen ins Englische:\" (Translate the following sentences from German to English). By utilizing language-specific instructions, there is a semantic consistency established between the instruction and the source language, resulting in clearer and more accurate guidance. As shown in Figure 3."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.713,
+                0.885,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Supervised Instruction Fine-Tuning Supervised Instruction Fine-Tuning for machine translation tasks incorporates two pivotal aspects. Firstly, akin to the earlier phase of Continual Pre-training, we employ LoRA(Hu et al., 2021) to finely tune specific parameters of Language Learning Models (LLMs), thereby enhancing their efficiency. LoRA(Hu et al., 2021) plays a crucial role in preventing model overfitting and leads to notable performance improvements. With this approach, we judiciously fine-tune a subset of model parameters using low-rank updates, striking a delicate balance between model adaptation and computational effi"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.486,
+                0.929,
+                0.517,
+                0.941
+            ],
+            "angle": 0,
+            "content": "642"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.086,
+                0.172,
+                0.1
+            ],
+            "angle": 0,
+            "content": "ciency."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.106,
+                0.49,
+                0.283
+            ],
+            "angle": 0,
+            "content": "Secondly, as emphasized in prior studies(Zhou et al., 2023; Maillard et al., 2023; Xu et al., 2023), LLMs exhibit benefits from a limited yet high-quality dataset. To ensure optimal data quality during the fine-tuning process, we leverage exceptional data sources. In line with previous research, we make use of meticulously curated human-written datasets derived from the WMT test data. These datasets undergo rigorous quality control measures, rendering them an ideal choice for fine-tuning purposes."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.31,
+                0.262,
+                0.327
+            ],
+            "angle": 0,
+            "content": "4 Experiments"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.347,
+                0.418,
+                0.362
+            ],
+            "angle": 0,
+            "content": "4.1 Datasets and Evaluation Metrics"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.377,
+                0.49,
+                0.569
+            ],
+            "angle": 0,
+            "content": "The overall data statistics are shown in Table 5 of Appendix A. For Stage 2, we utilized the WMT bilingual training dataset consisting of English \\(\\Leftrightarrow\\) German (En \\(\\Leftrightarrow\\) De) and English \\(\\Leftrightarrow\\) Chinese (En \\(\\Leftrightarrow\\) Zh) sentence pairs. The En \\(\\Leftrightarrow\\) De dataset comprised approximately 4.5 million pairs, while the En \\(\\Leftrightarrow\\) Zh dataset contained around 25 million pairs. Due to the higher number of En \\(\\Leftrightarrow\\) Zh pairs compared to En \\(\\Leftrightarrow\\) De, we sampled 4.5 million En \\(\\Leftrightarrow\\) Zh pairs for our experiments. Overall, the combined dataset contained nearly 1 billion tokens."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.575,
+                0.49,
+                0.688
+            ],
+            "angle": 0,
+            "content": "For Stage 3, we employed the newstest2017-2020 dataset for both \\(\\mathrm{En} \\Leftrightarrow \\mathrm{Zh}\\) and \\(\\mathrm{En} \\Leftrightarrow \\mathrm{De}\\) translation tasks. This dataset included a total of 37.6 thousand sentence pairs for each language direction. To ensure consistency across the source language and target language, we organize these sentence pairs into Source-Language Consistent Instructions."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.692,
+                0.49,
+                0.805
+            ],
+            "angle": 0,
+            "content": "We additionally incorporate the test sets from the WMT22 competition, which are carefully curated to include more recent content from diverse domains such as news, social media, e-commerce, and conversations. The test sets for the \\(\\mathrm{De} \\Rightarrow \\mathrm{En}\\), \\(\\mathrm{En} \\Rightarrow \\mathrm{De}\\), \\(\\mathrm{Zh} \\Rightarrow \\mathrm{En}\\), and \\(\\mathrm{En} \\Rightarrow \\mathrm{Zh}\\) tasks consist of 1984, 2037, 1875, and 2037 samples, respectively."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.809,
+                0.491,
+                0.921
+            ],
+            "angle": 0,
+            "content": "For automatic evaluation, we utilize Sacre-BLEU, which implements BLEU(Papineni et al., 2002), and COMET(Rei et al., 2020) from Unbabel/wmt22-comet-da. SacreBLEU calculates similarity based on n-gram matching, while COMET leverages cross-lingual pretrained models for evaluation."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.085,
+                0.603,
+                0.1
+            ],
+            "angle": 0,
+            "content": "4.2 Setup"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.108,
+                0.885,
+                0.219
+            ],
+            "angle": 0,
+            "content": "We conducted our experiments using HuggingFace Transformers with open-source LLMs from the LLaMA(Touvron et al., 2023) family. Specifically, we utilized LLaMA2-7b with matched parameters as our foundation model. Additionally, we included LLaMA2-13b to explore the impact of different model sizes."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.222,
+                0.886,
+                0.399
+            ],
+            "angle": 0,
+            "content": "Due to computational constraints, we did not reproduce the foundational experiments from Stage 1. After Stage 1, we selected Chinese-LLaMA2(Cui et al., 2023) as our new foundation model. Chinese-LLaMA2 is an extended and optimized version of Llama-2, specifically tailored for Chinese language understanding and instruction comprehension. It incorporates a larger Chinese vocabulary and undergoes incremental pretraining on a large-scale Chinese dataset, which further enhances its semantic understanding capabilities."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.4,
+                0.884,
+                0.496
+            ],
+            "angle": 0,
+            "content": "For Stage 2, Continual Pre-training, and Stage 3, Supervised Fine-Tuning, we referred to the hyperparameters employed in the Chinese-LLaMA2 project. During Stage 2, we trained the model for 1 epoch, and for Stage 3, we extended the training to 3 epochs."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.498,
+                0.884,
+                0.561
+            ],
+            "angle": 0,
+            "content": "Our experiments were conducted on 8 Nvidia GPUs with 64GB of memory each, utilizing DeepSpeed(Rasley et al., 2020) ZeRO 2 for model parallelization."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.576,
+                0.63,
+                0.59
+            ],
+            "angle": 0,
+            "content": "4.3 Baselines"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.599,
+                0.884,
+                0.807
+            ],
+            "angle": 0,
+            "content": "We evaluate our method against two baseline categories, consistent with previous studies. Firstly, we compare our approach to prior studies that share our goal of leveraging LLMs for translation. Secondly, we benchmark against the current state-of-the-art (SoTA) translation models. It's important to note that this comparison may not be entirely fair due to disparities in training data and model architectures. For example, there is a significant contrast between the 175B GPT-3.5 model and our 7B model. Nevertheless, by utilizing the same test set, we can gain insights into our model's current performance."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.81,
+                0.885,
+                0.922
+            ],
+            "angle": 0,
+            "content": "In the category of prior similar work, we compare our model to the following approaches: BigTranslate(Yang et al., 2023), which extends LLaMA-1-13B to cover over 100 translation directions; TIM(Zeng et al., 2023), which leverages correct and incorrect examples to aid LLMs in learning translation; ParroT(Jiao et al., 2023), through"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.486,
+                0.929,
+                0.516,
+                0.941
+            ],
+            "angle": 0,
+            "content": "643"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.127,
+                0.082,
+                0.871,
+                0.433
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Models</td><td colspan=\"2\">De→En</td><td colspan=\"2\">En→De</td><td colspan=\"2\">Zh→En</td><td colspan=\"2\">En→Zh</td></tr><tr><td>BLEU</td><td>COMET</td><td>BLEU</td><td>COMET</td><td>BLEU</td><td>COMET</td><td>BLEU</td><td>COMET</td></tr><tr><td colspan=\"9\">SoTA models</td></tr><tr><td>NLLB-54B(Team et al., 2022)</td><td>26.89</td><td>78.94</td><td>34.50</td><td>86.45</td><td>16.56</td><td>70.70</td><td>27.38</td><td>78.91</td></tr><tr><td>NLLB-54B Fine-tune</td><td>27.34</td><td>79.86</td><td>35.07</td><td>86.95</td><td>17.26</td><td>71.35</td><td>27.89</td><td>80.13</td></tr><tr><td>GPT-3.5-D, zero-shot</td><td>30.90</td><td>84.79</td><td>31.80</td><td>85.61</td><td>25.00</td><td>81.60</td><td>38.30</td><td>85.76</td></tr><tr><td>GPT-3.5-T, zero-shot</td><td>33.10</td><td>85.50</td><td>34.40</td><td>87.00</td><td>26.60</td><td>82.90</td><td>44.90</td><td>87.00</td></tr><tr><td>GPT-4</td><td>33.87</td><td>85.62</td><td>35.38</td><td>87.44</td><td>27.20</td><td>82.79</td><td>43.98</td><td>87.49</td></tr><tr><td colspan=\"9\">Prior Similar Studies</td></tr><tr><td>TIM-7B(Zeng et al., 2023)</td><td>27.91</td><td>82.80</td><td>25.59</td><td>82.56</td><td>19.33</td><td>75.46</td><td>19.33</td><td>75.46</td></tr><tr><td>Parrot-7B(Jiao et al., 2023)</td><td>29.80</td><td>83.00</td><td>26.10</td><td>81.60</td><td>20.20</td><td>75.90</td><td>30.30</td><td>80.30</td></tr><tr><td>SWIE-7B(Chen et al., 2023)</td><td>30.48</td><td>82.97</td><td>27.21</td><td>82.36</td><td>21.30</td><td>76.48</td><td>31.24</td><td>80.63</td></tr><tr><td>ALMA-7B(Xu et al., 2023)</td><td>29.56</td><td>83.95</td><td>30.31</td><td>85.59</td><td>23.64</td><td>79.78</td><td>36.48</td><td>85.05</td></tr><tr><td>Parrot-13B(Jiao et al., 2023)</td><td>31.10</td><td>83.60</td><td>28.10</td><td>82.60</td><td>21.70</td><td>76.70</td><td>31.70</td><td>81.00</td></tr><tr><td>BigTranslate-13B(Yang et al., 2023)</td><td>23.35</td><td>80.68</td><td>21.48</td><td>78.81</td><td>14.16</td><td>74.26</td><td>28.56</td><td>81.31</td></tr><tr><td>Bayling-13B(Zhang et al., 2023)</td><td>27.34</td><td>83.02</td><td>25.62</td><td>82.69</td><td>20.12</td><td>77.72</td><td>37.92</td><td>84.62</td></tr><tr><td>ALMA-13B(Xu et al., 2023)</td><td>31.14</td><td>84.56</td><td>31.47</td><td>85.62</td><td>25.46</td><td>80.21</td><td>39.84</td><td>85.96</td></tr><tr><td>Ours</td><td colspan=\"8\">Our Recipe with Backbone Model: LLaMA2(Touvron et al., 2023)</td></tr><tr><td>7B Stage3</td><td>30.02</td><td>84.09</td><td>25.40</td><td>82.30</td><td>20.59</td><td>76.18</td><td>30.60</td><td>80.40</td></tr><tr><td>7B Stage1,3*</td><td>25.20</td><td>78.32</td><td>12.50</td><td>69.19</td><td>20.90</td><td>76.40</td><td>35.00</td><td>84.32</td></tr><tr><td>7B Stage2,3</td><td>31.14</td><td>84.70</td><td>30.50</td><td>85.62</td><td>21.97</td><td>78.45</td><td>39.00</td><td>85.79</td></tr><tr><td>7B Stage1,2,3*</td><td>30.10</td><td>83.96</td><td>29.90</td><td>83.86</td><td>22.20</td><td>79.88</td><td>41.10</td><td>86.37</td></tr><tr><td>13B Stage3</td><td>31.70</td><td>84.39</td><td>28.80</td><td>83.87</td><td>21.40</td><td>77.68</td><td>35.90</td><td>84.23</td></tr><tr><td>13B Stage1,3*</td><td>26.13</td><td>78.65</td><td>12.79</td><td>72.23</td><td>21.40</td><td>78.28</td><td>37.34</td><td>85.27</td></tr><tr><td>13B Stage2,3</td><td>32.24</td><td>85.17</td><td>32.53</td><td>86.14</td><td>22.57</td><td>79.05</td><td>40.40</td><td>85.98</td></tr><tr><td>13B Stage1,2,3*</td><td>30.21</td><td>84.26</td><td>30.41</td><td>84.72</td><td>23.10</td><td>80.53</td><td>42.30</td><td>86.65</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.113,
+                0.443,
+                0.883,
+                0.501
+            ],
+            "angle": 0,
+            "content": "Table 1: The overall results. Note: * Due to computational constraints, we did not reproduce the foundational experiments from Stage 1, but instead directly utilized the Chinese-Llama2(Cui et al., 2023) that had undergone similar training. Since Chinese-Llama2(Cui et al., 2023) was only trained in Chinese during Stage 1, our main analysis about Stage 1 focuses on its performance in \\( \\mathrm{{Zh}} \\Rightarrow  \\mathrm{{En}} \\) and \\( \\mathrm{{En}} \\Rightarrow  \\mathrm{{Zh}} \\) translations."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.525,
+                0.49,
+                0.703
+            ],
+            "angle": 0,
+            "content": "three types of instructions including translation instruction, contrastive instruction, and error-guided instruction, improves the translation performance of LLM after SFT; SWIE(Chen et al., 2023), which enhances LLMs in translation through instruction augmentation; BayLing(Zhang et al., 2023), which incorporates interactive translation instructions; and ALMA(Xu et al., 2023), a two-stage finetuning method that initially fine-tunes on monolingual data and subsequently on a small set of high-quality parallel data."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.713,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": "In the SoTA models category, we consider the following: the NLLB-54B(Team et al., 2022) model, the largest and best translation model released in the NLLB family; the zero-shot performance of GPT3.5-text-davinci-003 (GPT-3.5-D) and GPT-3.5-turbo-0301 (GPT-3.5-T). Additionally, we present the zero-shot results for GPT-4. For a fair comparison, we also compared the results of fine-tuning NLLB-54B model with 37.6k data in Stage 3. To evaluate these baselines, we employ the same test data and evaluation metrics, reporting BLEU(Papineni et al., 2002) and COMET(Rei et al., 2020) scores as provided in their respective"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.51,
+                0.527,
+                0.567,
+                0.541
+            ],
+            "angle": 0,
+            "content": "papers."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.555,
+                0.724,
+                0.571
+            ],
+            "angle": 0,
+            "content": "5 Results and Analysis"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.582,
+                0.885,
+                0.727
+            ],
+            "angle": 0,
+            "content": "As shown in Table 1, overall, our results outperform all previous studies, NLLB-54B(Team et al., 2022), and GPT-3.5-D, except for a slight lag in \\(\\mathrm{Zh}\\Rightarrow \\mathrm{En}\\). Even our 7B model surpasses the results of other works. Particularly in the \\(\\mathrm{En}\\Rightarrow \\mathrm{Zh}\\) direction, our BLEU score is approximately 2.5 higher than the previous state-of-the-art. These findings are a testament to the effectiveness of our approach."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.739,
+                0.803,
+                0.755
+            ],
+            "angle": 0,
+            "content": "5.1 Assessing the Impact of Stage 1"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.76,
+                0.885,
+                0.887
+            ],
+            "angle": 0,
+            "content": "Just as mentioned earlier, we didn't specifically train Llama2 in Stage 1, but instead directly utilized the Chinese-Llama2(Cui et al., 2023) that had undergone similar training. Since Chinese-Llama2(Cui et al., 2023) was only trained in Chinese during Stage 1, our main analysis focuses on its performance in \\(\\mathrm{Zh}\\Rightarrow \\mathrm{En}\\) and \\(\\mathrm{En}\\Rightarrow \\mathrm{Zh}\\) translations."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.89,
+                0.885,
+                0.921
+            ],
+            "angle": 0,
+            "content": "As shown in Table 1, our findings align with previous research conclusions that incremental train"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.486,
+                0.929,
+                0.516,
+                0.941
+            ],
+            "angle": 0,
+            "content": "644"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.49,
+                0.246
+            ],
+            "angle": 0,
+            "content": "ing on monolingual data is beneficial. Furthermore, we discovered that this benefit primarily affects the target language in translation tasks. For example, we observed a significant improvement in the performance of the 7B model on the \\(\\mathrm{En} \\Rightarrow \\mathrm{Zh}\\) test set, where the BLEU score increased from 30.60 to 35.00, a substantial improvement of 4.4 points. However, the improvement in the \\(\\mathrm{Zh} \\Rightarrow \\mathrm{En}\\) direction was limited, indicating that the role of Stage 1 is to enhance generation rather than comprehension."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.246,
+                0.49,
+                0.358
+            ],
+            "angle": 0,
+            "content": "Additionally, we found that performing incremental training on only one monolingual dataset had disastrous effects on translation tasks in other languages. For example, on the \\(\\mathrm{En} \\Rightarrow \\mathrm{De}\\) test set, the BLEU score plummeted from 25.40 to 12.50. Therefore, for multilingual translation, it is crucial to conduct Stage 1 training on multiple languages."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.37,
+                0.465,
+                0.385
+            ],
+            "angle": 0,
+            "content": "5.2 Measuring the Effectiveness of Stage 2"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.39,
+                0.49,
+                0.616
+            ],
+            "angle": 0,
+            "content": "As shown in Table 1, Llama2(Touvron et al., 2023) demonstrates improved quality across various test sets after Stage 2 training. An interesting observation, considering Llama2 as a large model primarily focused on English, is that the enhancement in English-Other translations is particularly noteworthy after Stage 2 Training. For instance, the 7B model exhibits remarkable improvements in \\(\\mathrm{En}\\Rightarrow \\mathrm{De}\\), with the BLEU score increasing from 25.40 to 30.50, and in \\(\\mathrm{En}\\Rightarrow \\mathrm{Zh}\\), where it rises from 30.60 to 39.00. The magnitude of these improvements is quite significant. Encouragingly, there are also improvements observed in translations from other languages to English."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.616,
+                0.49,
+                0.793
+            ],
+            "angle": 0,
+            "content": "An even more intriguing finding is that, as mentioned before, since Chinese-Llama2(Cui et al., 2023) only underwent Stage 1 training on Chinese, its translation performance substantially deteriorates in the \\(\\mathrm{En} \\Rightarrow \\mathrm{De}\\) direction. However, with the magical touch of Stage 2 training, these capabilities are miraculously restored. The 7B model, on \\(\\mathrm{En} \\Rightarrow \\mathrm{De}\\), rebounds from 12.50 to 29.90, approaching the results of the original Llama2(Touvron et al., 2023). These outcomes effectively affirm the effectiveness of Stage 2."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.794,
+                0.49,
+                0.921
+            ],
+            "angle": 0,
+            "content": "After considering the overall process, we are interested in understanding the impact of Stage 2 only. As mentioned before, LLMs typically include two main types of models: Foundation Models and Chat Models. Evaluating Stage 2 essentially assesses the Foundation Model by using an n-shot evaluation, which includes both zero-shot and 5-shot evaluations. We have noticed that zero-shot"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.885,
+                0.133
+            ],
+            "angle": 0,
+            "content": "evaluations can occur hallucinations. Hence, we are presenting the results of the 5-shot evaluation in Table 2."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.526,
+                0.142,
+                0.875,
+                0.203
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Models</td><td colspan=\"2\">Zh→En</td><td colspan=\"2\">En→Zh</td></tr><tr><td></td><td>BLEU</td><td>COMET</td><td>BLEU</td><td>COMET</td></tr><tr><td>Baseline</td><td>20.63</td><td>76.32</td><td>29.96</td><td>79.34</td></tr><tr><td>+ Stage2</td><td>21.64</td><td>78.07</td><td>38.62</td><td>85.30</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.508,
+                0.213,
+                0.885,
+                0.241
+            ],
+            "angle": 0,
+            "content": "Table 2: Results of the five-shot results based on Llama2-7B(Touvron et al., 2023) model."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.274,
+                0.831,
+                0.29
+            ],
+            "angle": 0,
+            "content": "5.3 Analyzing the Outcomes of Stage 3"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.295,
+                0.885,
+                0.456
+            ],
+            "angle": 0,
+            "content": "To evaluate the effectiveness of Source-Language Consistent Instruction in Stage 3, we conducted a comparative experiment using English-Fixed Instruction. The results of the experiment are presented in Table 3. It is evident that in the \\(\\mathrm{En}\\Rightarrow \\mathrm{De}\\) and \\(\\mathrm{De}\\Rightarrow \\mathrm{En}\\) directions, the performance of these two types of instructions is quite similar. However, in the \\(\\mathrm{Zh}\\Rightarrow \\mathrm{En}\\) and \\(\\mathrm{En}\\Rightarrow \\mathrm{Zh}\\) directions, the use of Source-Language Consistent Instruction clearly outperforms."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.518,
+                0.467,
+                0.88,
+                0.545
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Models</td><td>De⇒En</td><td>En⇒De</td><td>Zh⇒En</td><td>En⇒Zh</td></tr><tr><td>Stage 3</td><td>30.02</td><td>25.40</td><td>20.59</td><td>30.60</td></tr><tr><td>w/o</td><td>30.40</td><td>25.20</td><td>18.39</td><td>28.30</td></tr><tr><td>Stage2,3</td><td>31.14</td><td>30.50</td><td>21.91</td><td>39.00</td></tr><tr><td>w/o</td><td>31.00</td><td>30.23</td><td>18.93</td><td>38.69</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.508,
+                0.553,
+                0.883,
+                0.597
+            ],
+            "angle": 0,
+            "content": "Table 3: Results of the comparative experiments based on Llama2-7B(Touvron et al., 2023) model. [w/o] means using English-Fixed Instruction."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.614,
+                0.885,
+                0.71
+            ],
+            "angle": 0,
+            "content": "We believe that the similarity between English and German, as they belong to the same language family, contributes to the lack of noticeable differences. However, when dealing with cross-language pairs, employing Source-Language Consistent Instruction further enhances the translation quality."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.508,
+                0.723,
+                0.862,
+                0.739
+            ],
+            "angle": 0,
+            "content": "5.4 Comparing the Difference with ALMA"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.745,
+                0.884,
+                0.825
+            ],
+            "angle": 0,
+            "content": "We have noticed that our work shares some similarities with ALMA(Xu et al., 2023) in terms of the process, involving Continual Pre-training followed by Supervised Fine-Tuning. However, there are notable differences between our approaches."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.826,
+                0.885,
+                0.922
+            ],
+            "angle": 0,
+            "content": "ALMA suggests that the impact of bilingual data is reduced in the era of large models. In contrast, we firmly believe in the continued strength of bilingual data and its application in Continual Pre-training through Interlinear Text Format Documents. While ALMA acknowledges the effective"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.486,
+                0.929,
+                0.516,
+                0.941
+            ],
+            "angle": 0,
+            "content": "645"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.49,
+                0.246
+            ],
+            "angle": 0,
+            "content": "ness of conducting Continual Pre-training on monolingual data, we have also validated this finding in our own work and reached the same conclusion. However, it is important to note that our approach primarily enhances the multilingual generation capability of the large model itself, rather than being specifically tailored to translation tasks. Furthermore, ALMA utilizes a significantly larger training dataset, ranging from 13B to 20B, compared to our own."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.267,
+                0.475,
+                0.318
+            ],
+            "angle": 0,
+            "content": "6 Ablation Study: What if we directly employ a large quantity of translation data for SFT?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.332,
+                0.49,
+                0.445
+            ],
+            "angle": 0,
+            "content": "Both Continual Pre-training and Supervised Fine-Tuning involve incremental training on the original model. However, if we skip Stage 2 training and directly utilize the translation data from Stage 2 as instruction data for SFT, i.e., conducting SFT directly with a substantial amount of translation data, will it yield consistent improvement?"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.124,
+                0.46,
+                0.485,
+                0.521
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Data Size</td><td>De→En</td><td>En→De</td><td>Zh→En</td><td>En→Zh</td></tr><tr><td>37.6K</td><td>30.02</td><td>25.40</td><td>20.59</td><td>30.60</td></tr><tr><td>400K</td><td>30.20</td><td>25.60</td><td>18.49</td><td>31.74</td></tr><tr><td>4,000K</td><td>30.66</td><td>25.12</td><td>20.77</td><td>32.22</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.114,
+                0.53,
+                0.489,
+                0.575
+            ],
+            "angle": 0,
+            "content": "Table 4: Results of the ablation experiments based on Llama2-7B(Touvron et al., 2023) model under different Stage 3 data size."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.595,
+                0.49,
+                0.708
+            ],
+            "angle": 0,
+            "content": "To address this question, we conducted an ablation experiment. Our Stage 3 training data consisted of \\(37.6\\mathrm{k}\\) samples. Randomly selecting and merging some data from the Stage 2 training data with the Stage 3 training data, we created three sets: \\(37.6\\mathrm{K},400\\mathrm{K}\\), and \\(4,000\\mathrm{K}\\). The experimental results are presented in Table 4."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.71,
+                0.49,
+                0.775
+            ],
+            "angle": 0,
+            "content": "We found that augmenting the training data in Stage 3 slightly improved translation quality for certain test sets. This indicates that a small amount of high-quality data is sufficient for the SFT stage."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.777,
+                0.491,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Now, our focus is solely on the translation task. However, if we were conducting multi-task SFT, it is unlikely that other tasks would have as extensive data as machine translation. Therefore, using a large amount of translation data during SFT would result in the problem of imbalanced data distribution with other tasks. Hence, the optimal approach would still be to utilize this substantial data during the earlier stage of Continual Pre-training."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.084,
+                0.651,
+                0.1
+            ],
+            "angle": 0,
+            "content": "7 Conclusions"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.111,
+                0.887,
+                0.562
+            ],
+            "angle": 0,
+            "content": "In this study, we have introduced a novel paradigm for enhancing the translation capabilities of large language models in machine translation tasks. Our three-stage approach, including Secondary Pretraining using Extensive Monolingual Data, Continual Pre-training with Interlinear Text Format Documents, and Leveraging Source-Language Consistent Instruction for Supervised Fine-Tuning, addresses the limitations of previous strategies and offers notable improvements in translation quality. We emphasize the significance of pre-training stages in enhancing LLMs' cross-lingual alignment abilities and the effectiveness of using a smaller but high-quality set of bilingual data during supervised fine-tuning. Notably, Stage2, which involves Continual Pre-training with Interlinear Text Format Documents, stands out as a highly efficient method, requiring minimal training data. Furthermore, aligning the instructional setting with the source language during supervised fine-tuning, as observed in Stage3, yields positive effects. The findings from this paper contribute to advancing the field of machine translation and offer valuable insights for optimizing the translation capabilities of large language models. Future research can explore additional language pairs, alternative data augmentation techniques, and different pre-training strategies to further refine our proposed paradigm."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.574,
+                0.646,
+                0.59
+            ],
+            "angle": 0,
+            "content": "8 Limitations"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.601,
+                0.886,
+                0.842
+            ],
+            "angle": 0,
+            "content": "Despite notable contributions, this study has certain limitations. Firstly, the proposed method exhibits slightly reduced performance in the \\(\\mathrm{Zh} \\Rightarrow \\mathrm{En}\\) translation direction, necessitating further analysis and improvements. Secondly, the presence of illusionary translations within large models was observed but not extensively explored. Future research should delve deeper into this phenomenon. Lastly, while the paper primarily focuses on SFT for machine translation, opportunities exist to explore SFT techniques in diverse contexts such as style translation and colloquial translation. Addressing these limitations would further enhance the effectiveness and applicability of the proposed methods."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.87,
+                0.611,
+                0.886
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.51,
+                0.894,
+                0.884,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.486,
+                0.929,
+                0.517,
+                0.941
+            ],
+            "angle": 0,
+            "content": "646"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.135,
+                0.086,
+                0.49,
+                0.216
+            ],
+            "angle": 0,
+            "content": "Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. CoRR, abs/2005.14165."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.118,
+                0.227,
+                0.487,
+                0.28
+            ],
+            "angle": 0,
+            "content": "Yijie Chen, Yijin Liu, Fandong Meng, Yufeng Chen, Jinan Xu, and Jie Zhou. 2023. Improving translation faithfulness of large language models via augmenting instructions."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.118,
+                0.29,
+                0.489,
+                0.602
+            ],
+            "angle": 0,
+            "content": "Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. 2023. Palm: Scaling language modeling with pathways. J. Mach. Learn. Res., 24:240:1-240:113."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.614,
+                0.487,
+                0.654
+            ],
+            "angle": 0,
+            "content": "Yiming Cui, Ziqing Yang, and Xin Yao. 2023. Efficient and effective text encoding for chinese llama and alpaca. arXiv preprint arXiv:2304.08177."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.664,
+                0.488,
+                0.782
+            ],
+            "angle": 0,
+            "content": "Marjan Ghazvininejad, Omer Levy, Yinhan Liu, and Luke Zettlemoyer. 2019. Mask-predict: Parallel decoding of conditional masked language models. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, November 3-7, 2019, pages 6111-6120. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.792,
+                0.488,
+                0.87
+            ],
+            "angle": 0,
+            "content": "Jiatao Gu, James Bradbury, Caiming Xiong, Victor O. K. Li, and Richard Socher. 2018. Non-autoregressive neural machine translation. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.881,
+                0.488,
+                0.921
+            ],
+            "angle": 0,
+            "content": "Jiaxin Guo, Minghan Wang, Daimeng Wei, Hengchao Shang, Yuxia Wang, Zongyao Li, Zhengzhe Yu, Zhanglin Wu, Yimeng Chen, Chang Su, Min Zhang,"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.921
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.529,
+                0.086,
+                0.882,
+                0.125
+            ],
+            "angle": 0,
+            "content": "Lizhi Lei, Shimin Tao, and Hao Yang. 2021. Self-distillation mixup training for non-autoregressive neural machine translation. CoRR, abs/2112.11640."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.136,
+                0.883,
+                0.213
+            ],
+            "angle": 0,
+            "content": "Amr Hendy, Mohamed Abdelrehim, Amr Sharaf, Vikas Raunak, Mohamed Gabr, Hitokazu Matsushita, Young Jin Kim, Mohamed Affify, and Hany Hassan Awadalla. 2023. How good are GPT models at machine translation? A comprehensive evaluation. CoRR, abs/2302.09210."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.224,
+                0.882,
+                0.276
+            ],
+            "angle": 0,
+            "content": "Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.286,
+                0.883,
+                0.351
+            ],
+            "angle": 0,
+            "content": "Wenxiang Jiao, Jen tse Huang, Wenxuan Wang, Zhiwei He, Tian Liang, Xing Wang, Shuming Shi, and Zhaopeng Tu. 2023. Parrot: Translating during chat using large language models tuned with human translation and feedback."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.362,
+                0.883,
+                0.492
+            ],
+            "angle": 0,
+            "content": "Jean Maillard, Cynthia Gao, Elahe Kalbassi, Kaushik Ram Sadagopan, Vedanuj Goswami, Philipp Koehn, Angela Fan, and Francisco Guzmán. 2023. Small data, big impact: Leveraging minimal data for effective machine translation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2023, Toronto, Canada, July 9-14, 2023, pages 2740-2756. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.502,
+                0.883,
+                0.58
+            ],
+            "angle": 0,
+            "content": "Kishore Papineni, Salim Roukos, Todd Ward, and Wei Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, July 6-12, 2002, Philadelphia, PA, USA, pages 311-318. ACL."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.59,
+                0.883,
+                0.682
+            ],
+            "angle": 0,
+            "content": "Jeff Rasley, Samyam Rajbhandari, Olatunjri Ruwase, and Yuxiong He. 2020. Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In KDD '20: The 26th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Virtual Event, CA, USA, August 23-27, 2020, pages 3505-3506. ACM."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.691,
+                0.883,
+                0.783
+            ],
+            "angle": 0,
+            "content": "Ricardo Rei, Craig Stewart, Ana C. Farinha, and Alon Lavie. 2020. COMET: A neural framework for MT evaluation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020, pages 2685-2702. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.793,
+                0.883,
+                0.833
+            ],
+            "angle": 0,
+            "content": "Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Improving neural machine translation models with monolingual data."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.842,
+                0.883,
+                0.921
+            ],
+            "angle": 0,
+            "content": "NLLB Team, Marta R. Costa-jussa, James Cross, Onur Celebi, Maha Elbayad, Kenneth Heafield, Kevin Hefernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, Anna Sun, Skyler Wang, Guillaume Wenzek, Al Youngblood, Bapi Akula, Loic Barrault, Gabriel Mejia Gonzalez, Prangthip Hansanti,"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.512,
+                0.086,
+                0.883,
+                0.921
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.486,
+                0.929,
+                0.516,
+                0.94
+            ],
+            "angle": 0,
+            "content": "647"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.132,
+                0.086,
+                0.49,
+                0.204
+            ],
+            "angle": 0,
+            "content": "John Hoffman, Semarley Jarrett, Kaushik Ram Sadagopan, Dirk Rowe, Shannon Spruit, Chau Tran, Pierre Andrews, Necip Fazil Ayan, Shruti Bhosale, Sergey Edunov, Angela Fan, Cynthia Gao, Vedanuj Goswami, Francisco Guzmán, Philipp Koehn, Alexandre Mourachko, Christophe Ropers, Safiyyah Saleem, Holger Schwenk, and Jeff Wang. 2022. No language left behind: Scaling human-centered machine translation."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.116,
+                0.214,
+                0.49,
+                0.514
+            ],
+            "angle": 0,
+            "content": "Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. 2023. Llama 2: Open foundation and finetuned chat models."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.116,
+                0.524,
+                0.49,
+                0.618
+            ],
+            "angle": 0,
+            "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5998-6008."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.116,
+                0.627,
+                0.49,
+                0.732
+            ],
+            "angle": 0,
+            "content": "Minghan Wang, Jiaxin Guo, Yuxia Wang, Yimeng Chen, Chang Su, Daimeng Wei, Min Zhang, Shimin Tao, and Hao Yang. 2021. HI-CMLM: improve CMLM with hybrid decoder input. In Proceedings of the 14th International Conference on Natural Language Generation, INLG 2021, Aberdeen, Scotland, UK, 20-24 September, 2021, pages 167-171. Association for Computational Linguistics."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.116,
+                0.742,
+                0.49,
+                0.794
+            ],
+            "angle": 0,
+            "content": "Daimeng Wei, Zhanglin Wu, Hengchao Shang, Zongyao Li, Minghan Wang, Jiaxin Guo, Xiaoyu Chen, Zhengzhe Yu, and Hao Yang. 2023. Text style transfer back-translation."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.116,
+                0.805,
+                0.49,
+                0.858
+            ],
+            "angle": 0,
+            "content": "Haoran Xu, Young Jin Kim, Amr Sharaf, and Hany Hassan Awadalla. 2023. A paradigm shift in machine translation: Boosting translation performance of large language models."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.116,
+                0.868,
+                0.49,
+                0.921
+            ],
+            "angle": 0,
+            "content": "Wen Yang, Chong Li, Jiajun Zhang, and Chengqing Zong. 2023. Bigtranslate: Augmenting large language models with multilingual translation capability over 100 languages."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.51,
+                0.086,
+                0.885,
+                0.165
+            ],
+            "angle": 0,
+            "content": "Zhengzhe Yu, Jiaxin Guo, Minghan Wang, Daimeng Wei, Hengchao Shang, Zongyao Li, Zhanglin Wu, Yuxia Wang, Yimeng Chen, Chang Su, Min Zhang, Lizhi Lei, Shimin Tao, and Hao Yang. 2021. Jointtraining on symbiosis networks for deep nerval machine translation models. CoRR, abs/2112.11642."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.51,
+                0.175,
+                0.885,
+                0.215
+            ],
+            "angle": 0,
+            "content": "Jiali Zeng, Fandong Meng, Yongjing Yin, and Jie Zhou. 2023. Tim: Teaching large language models to translate with comparison."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.51,
+                0.224,
+                0.885,
+                0.304
+            ],
+            "angle": 0,
+            "content": "Shaolei Zhang, Qingkai Fang, Zhuocheng Zhang, Zhengrui Ma, Yan Zhou, Langlin Huang, Mengyu Bu, Shangtong Gui, Yunji Chen, Xilin Chen, and Yang Feng. 2023. Bayling: Bridging cross-lingual alignment and instruction following through interactive translation for large language models."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.51,
+                0.313,
+                0.885,
+                0.38
+            ],
+            "angle": 0,
+            "content": "Chunting Zhou, Pengfei Liu, Puxin Xu, Srini Iyer, Jiao Sun, Yuning Mao, Xuezhe Ma, Avia Efrat, Ping Yu, Lili Yu, Susan Zhang, Gargi Ghosh, Mike Lewis, Luke Zettlemoyer, and Omer Levy. 2023. LIMA: less is more for alignment. CoRR, abs/2305.11206."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.404,
+                0.788,
+                0.42
+            ],
+            "angle": 0,
+            "content": "A Appendix A: Data Statistics"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.43,
+                0.885,
+                0.572
+            ],
+            "angle": 0,
+            "content": "Table 5 displays the comprehensive data statistics. For Stage 2, we utilized the WMT bilingual training dataset that includes English \\(\\Leftrightarrow\\) German (En \\(\\Leftrightarrow\\) De) and English \\(\\Leftrightarrow\\) Chinese (En \\(\\Leftrightarrow\\) Zh) sentence pairs. The En \\(\\Leftrightarrow\\) De dataset comprised approximately 4.5 million pairs, while for the En \\(\\Leftrightarrow\\) Zh dataset, we randomly sampled an equivalent number of pairs from the total 25 million pairs. In total, the combined dataset contained close to 1B tokens."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.574,
+                0.885,
+                0.701
+            ],
+            "angle": 0,
+            "content": "Moving on to Stage 3, we utilized the newstest2017-2020 dataset for both the \\(\\mathrm{En} \\Leftrightarrow \\mathrm{De}\\) and \\(\\mathrm{En} \\Leftrightarrow \\mathrm{Zh}\\) translation tasks. This dataset comprised 37.6 thousand sentence pairs for each language direction. To maintain coherence between the source and target languages, we categorized these sentence pairs into Source-Language Consistent Instructions."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.714,
+                0.85,
+                0.748
+            ],
+            "angle": 0,
+            "content": "B Appendix B: Stage 2 Training Data Sample"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.757,
+                0.882,
+                0.787
+            ],
+            "angle": 0,
+            "content": "Figure 4 displays samples of the Stage 2 training data."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.486,
+                0.929,
+                0.516,
+                0.94
+            ],
+            "angle": 0,
+            "content": "648"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.151,
+                0.187,
+                0.849,
+                0.266
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Stage</td><td>Data</td><td>Description</td></tr><tr><td>Stage 1</td><td>120G text</td><td>120G Chinese text described in Chinese-LLaMA2(Cui et al., 2023)</td></tr><tr><td>Stage 2</td><td>1B tokens</td><td>1B tokens including four directions: En↔De and En↔Zh. Each direction includes 4.5 million pairs.</td></tr><tr><td>Stage 3</td><td>37.6k pairs</td><td>37.6k pairs combined wmt newstest2017-2020 testset with all four directions.</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.373,
+                0.275,
+                0.625,
+                0.289
+            ],
+            "angle": 0,
+            "content": "Table 5: The overall data statistics."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.298,
+                0.516,
+                0.697,
+                0.779
+            ],
+            "angle": 0,
+            "content": "<s> But buses are faster and more frequent. \\n Aber Buses sind Schneller und haufter. </s> <s> 因此, 公布的历史成本不反映非消耗性财产充分而准确的估价; \\n Accordingly, the historical costs disclosed did not reflect the full and accurate valuation of non-expendable property; </s> Wir haben heute dem Bericht Ferber zugestimmt, obwohl er eine Grundlage für den Kauf der Parlamentsgebäude in Straßburg bildet. Unsere Zustimmung stellt aber keine Vorfestlegung auf den Sitz des Europäischen Parlaments dar. \\n We voted in favour of the Ferber report today, even though it forms a basis for the purchase of the Parliament building in Strasbourg. However, our approval does not constitute a prior decision regarding the seat of the European Parliament. </s> <s> By introducing greater flexibility in how the Fund is used and by reducing the number of redundancies from 1000 to 500, it will become an ever more effective instrument for helping to tackle the effects of the economic down-turn. \\n Durch die Einführung von mehr Flexibilität in der Nutzung der Fonds und durch die Reduzierung der Zahl der Entlassungen von 1000 auf 500 wird er ein noch effizienteres Instrument zur Bekämpfung der Folgen des Wirtschaftssabschwungs werden. </s> <s> During the period from 1 January to 30 June 2010, contributions totalling $19,299.39 were received from Angola ($10,000) and Congo ($9,299.39). \\n 2010年1月1日至6月30日期间,从安哥拉(10000美元)和刚果(9299.39美元)共收到19299.39美元捐款。 </s> <s> ..."
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.362,
+                0.797,
+                0.637,
+                0.813
+            ],
+            "angle": 0,
+            "content": "Figure 4: Stage 2 Training Data Sample"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.486,
+                0.929,
+                0.517,
+                0.941
+            ],
+            "angle": 0,
+            "content": "649"
+        }
+    ]
+]

2024/A Novel Paradigm Boosting Translation Capabilities of Large Language Models/8eb804a5-0f10-4fdb-af4a-6a552f7a4c4e_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:78bb3738963ae778481640eabd1e01fe6f0888d75c4ca658ff5618a1d6fd2273
+size 391416

2024/A Novel Paradigm Boosting Translation Capabilities of Large Language Models/full.md

@@ -0,0 +1,284 @@
+# A Novel Paradigm Boosting Translation Capabilities of Large Language Models
+
+Jiaxin Guo*, Hao Yang†, Zongyao Li
+
+Daimeng Wei, Hengchao Shang, Xiaoyu Chen
+
+{jiaxinguo1,yanghao30,lizongyao}  $@$  huawei.com
+
+{weidaimeng,shanghengchao,chenxiaoyu35}@huawei.com
+
+Huawei Translation Services Center, Beijing, China
+
+# Abstract
+
+This paper presents a study on strategies to enhance the translation capabilities of large language models (LLMs) in the context of machine translation (MT) tasks. The paper proposes a novel paradigm consisting of three stages: Secondary Pre-training using Extensive Monolingual Data, Continual Pre-training with Interlinear Text Format Documents, and Leveraging Source-Language Consistent Instruction for Supervised Fine-Tuning. Previous research on LLMs focused on various strategies for supervised fine-tuning (SFT), but their effectiveness has been limited. While traditional machine translation approaches rely on vast amounts of parallel bilingual data, our paradigm highlights the importance of using smaller sets of high-quality bilingual data. We argue that the focus should be on augmenting LLMs' cross-lingual alignment abilities during pre-training rather than solely relying on extensive bilingual data during SFT. Experimental results conducted using the Llama2(Touvron et al., 2023) model, particularly on Chinese-Llama2(Cui et al., 2023) after monolingual augmentation, demonstrate the improved translation capabilities of LLMs. A significant contribution of our approach lies in Stage2: Continual Pre-training with Interlinear Text Format Documents, which requires less than 1B training data, making our method highly efficient. Additionally, in Stage3, we observed that setting instructions consistent with the source language benefits the supervised fine-tuning process. Experimental results demonstrate that our approach surpasses previous work and achieves superior performance compared to models such as NLLB-54B(Team et al., 2022) and GPT3.5-text-davinci-003, despite having a significantly smaller parameter count of only 7B or 13B. This achievement establishes our method as a pioneering strategy in the field of machine translation.
+
+# 1 Introduction
+
+Translation capabilities of large language models (LLMs)(Brown et al., 2020; Chowdhery et al., 2023; Touvron et al., 2023) in machine translation (MT) tasks have been explored extensively in previous research(Jiao et al., 2023; Zeng et al., 2023; Chen et al., 2023; Xu et al., 2023; Yang et al., 2023; Zhang et al., 2023). However, achieving significant improvements in translation quality through supervised fine-tuning (SFT) strategies has proven challenging. Traditionally, machine translation relies on vast amounts of parallel bilingual data, but SFT only requires a small amount of high-quality bilingual data, highlighting a crucial distinction. It is a naive approach to consider using vast quantities of parallel bilingual data during SFT. However, experiments have shown that increasing the data volume yields limited improvements in quality and even leads to performance degradation on certain test sets. Thus, the question arises: are extensive parallel bilingual data useless in SFT, or are they being misused?
+
+In this paper, we propose a novel training paradigm, consisting of three stages, to boost the translation capabilities of LLMs. Our contributions include refining the training strategy for downstream tasks and emphasizing the enhancement of LLMs' cross-lingual alignment abilities during pretraining. These contributions address the limitations observed in previous research. Our training paradigm comprises the following stages:
+
+Stage 1: Continual Pre-training using Extensive Monolingual Data. Consistent with previous findings(Xu et al., 2023), we validate the effectiveness of monolingual data augmentation. Specifically, we perform SFT on a Chinese-Llama2(Cui et al., 2023) model, which undergoes monolingual data augmentation, thereby demonstrating improved translation capabilities.
+
+Stage 2: Continual Pre-training with
+
+Sentence-aligned Parallel Data. We construct interlinear text format from sentence-aligned bilingual parallel data and utilize them for continual pre-training of LLMs. Experimental results demonstrate the critical importance of this stage, resulting in a significant improvement in translation quality, particularly for English-Other translations. Stage 2 stands as a pivotal contribution in our paper, requiring less than 1B training data, thereby enhancing training efficiency.
+
+Stage 3: Leveraging Source-Language Consistent Instruction for Supervised Fine-Tuning. In SFT, we discover that using instruction aligned with the source language of the translation notably improves performance. Leveraging source-language consistent instructions during SFT yields significant enhancements.
+
+In summary, we introduce a three-stage training paradigm, highlighting the effectiveness of secondary pre-training, continual pre-training with interlinear text format documents, and leveraging source-language consistent instruction for supervised fine-tuning. These contributions address the limitations observed in previous research and pave the way for improved translation quality.
+
+# 2 Related Work
+
+# 2.1 Large Language Models
+
+Foundation Model Foundation Model, a product of pre-training, is a prominent type of Large Language Model. It has gained substantial recognition in recent years for its impressive capabilities in natural language processing tasks. The most prevalent architectural framework for such models is the Transformer, which employs a series of self-attention mechanisms to process input text efficiently.
+
+Among the state-of-the-art Large Language Models, notable examples include GPT-3(Brown et al., 2020) and Llama2(Touvron et al., 2023). These models have been widely lauded for their exceptional proficiency in understanding and generating natural language text. They showcase the remarkable potential of Foundation Models, pushing the boundaries of language processing and setting new benchmarks in various applications.
+
+Instruct/Chat Model Instruct/Chat Model, a variant of Large Language Models, is specifically developed through the process of Supervised Fine-Tuning (SFT). Unlike Foundation Models, which
+
+are pre-trained, Instruct/Chat Models undergo additional supervised training to enhance their performance in specific tasks such as instruction following or conversational dialogue.
+
+Supervised Fine-Tuning involves training the model on labeled datasets, where human annotators provide examples of desired input-output behavior. This approach enables Instruct/Chat Models to learn task-specific skills and exhibit improved performance in situations that require language understanding, generation, and interaction.
+
+Noteworthy advancements have been observed in Instruct/Chat Models, with notable examples including models like ChatGPT. These models have exhibited remarkable outcomes in conversational scenarios, demonstrating their potential in enabling interactive and engaging human-like conversations.
+
+# 2.2 Machine Translation Task
+
+Machine Translation Task refers to the process of automatically translating text from one language to another using computational methods.
+
+Traditional Methods Traditional machine translation methods primarily rely on encoder-decoder(Vaswani et al., 2017) models, where an encoder converts the source language sentence and a decoder produces the translated sentence. These methods heavily depend on large bilingual parallel corpora for training, aligning source sentences with their corresponding translations. Data augmentation(Sennrich et al., 2016; Wei et al., 2023) is a common practice in traditional machine translation. Some studies(Gu et al., 2018; Ghazvininejad et al., 2019; Wang et al., 2021; Guo et al., 2021; Yu et al., 2021) also investigate transforming them into parallel architectures to speed up inference efficiency.
+
+LLM-based Methods In recent years, Language Model (LM)-based approaches have gained attention in the field of machine translation. These approaches leverage pre-trained language models, such as the GPT (Generative Pre-trained Transformer)(Brown et al., 2020; Chowdhery et al., 2023; Touvron et al., 2023) series, and adapt them for translation tasks.
+
+One line of LLM-based methods focuses on zero-shot or few-shot translation by incorporating incontext learning(Hendy et al., 2023). By conditioning the LLM on a source sentence, the model can generate translations in the target language without explicitly using parallel data. This approach
+
+![](images/73f436e14355582f68c0076b18f2226258b42498f47b7b07c4ef7bfa9a5ff179.jpg)  
+Figure 1: The overall of our approach. Stage 1: Secondary Pre-training using Extensive Monolingual Data. Stage 2: Continual Pre-training with Interlinear Text Format Documents. Stage 3: Leveraging Source-Language Consistent Instruction for Supervised Fine-Tuning. *It should be noted that Stage 1 is considered non-essential.*
+
+![](images/6a3dc8f701c6c3d629cf4e86011d2f56bc3eea3b6069fec21035ccb49d0e3865.jpg)
+
+![](images/733b32e518371d86814b6b5a5aaa0d52670f85194e94c39dacc0e8f5f9cf944b.jpg)
+
+has shown promising results in enabling translation for language pairs with limited or no parallel resources.
+
+Another approach involves using a small amount of high-quality bilingual parallel data to construct translation-guiding instructions. These instructions explicitly define the translation behavior by providing source-language consistent cues during the supervised fine-tuning (SFT) process. By utilizing these specially crafted instructions, the LM can be fine-tuned to perform translation more accurately and robustly.
+
+Overall, LLM-based methods present alternative approaches to machine translation, exploring the potential of leveraging pre-trained models and incorporating limited parallel resources or high-quality instructions to enhance translation quality.
+
+# 3 A New Training Recipe
+
+We propose an innovative training strategy to enhance the translation capabilities of Large Language Models. As shown in Figure 1, our approach consists of three stages: (1) Secondary Pre-training using Extensive Monolingual Data, (2) Continual Pre-training with Interlinear Text Format Documents, and (3) Leveraging Source-Language Consistent Instruction for Supervised Fine-Tuning.
+
+# 3.1 Stage 1: Continual Pre-training using Extensive Monolingual Data
+
+In this stage, our aim is to enhance the training of large language models (LLMs) by utilizing di
+
+verse monolingual data. Currently, existing large models, such as Llama, are primarily pre-trained on English-centric corpora, resulting in relatively weaker comprehension and generation abilities in non-English languages. To expand the multilingual generation capabilities of LLMs, we suggest an incremental pre-training approach using extensive monolingual data.
+
+It is important to note that this stage primarily focuses on enhancing the intrinsic multilingual capacity of LLMs. While it is inherently related to machine translation tasks, it is not essential. On the one hand, we can select an existing LLM that already demonstrates robust multilingual capabilities as the base model for further training. On the other hand, even LLMs with limited multilingual support can benefit from the subsequent stages outlined in our methodology.
+
+# 3.2 Stage 2: Continual Pre-training with Sentence-aligned Parallel Data
+
+Interlinear Text Format Interlinear Text Format are a specific type of parallel text resource that consists of source sentences and their corresponding translations displayed in a aligned format. Each source sentence is accompanied by its translation, typically presented word-by-word or phrase-by-phrase, to facilitate a clear interlingual correspondence. We build the Sentence-aligned Parallel Data into this format. See Figure 2.
+
+Utilizing Interlinear Text Format offers several advantages for language understanding and trans
+
+![](images/024c4842a7908aaf690c6e1f062fab15f2309385c15e626385f7be37fe5c018b.jpg)  
+Figure 2: Interlinear Text Format Documents
+
+lation tasks. Firstly, these data provide explicit linguistic alignment at a fine-grained level, enabling the model to capture syntactic and semantic correspondences across languages. This aligns closely with the goals of machine translation, as it facilitates accurate encoding of source language information and improves the quality of generated translations. Additionally, interlinear data contributes to the learning of interlingual representations, allowing the model to better understand the relationship and transferability between languages.
+
+Continual Pre-training To leverage the benefits of Interlinear Text Documents, we propose a Continual Pre-training strategy based on the LoRA(Hu et al., 2021) (Low-Rank Adaptation of Large Language Models) framework. LoRA is a robust and effective pre-training approach for language models, introduced in recent research.
+
+By leveraging the inherent alignment information present in Interlinear Text Documents, the model learns to align and generate translations that maintain syntactic and semantic consistency with the source sentences. This continual training process allows the model to progressively improve its ability to capture cross-lingual correspondences, resulting in enhanced translation quality.
+
+# 3.3 Stage 3: Leveraging Source-Language Consistent Instruction for Supervised Fine-Tuning
+
+Source-Language Consistent Instruction In the field of machine translation, "Source-Language Consistent Instruction" refers to the practice of constructing translation instructions that maintain
+
+consistency with the source language, aiming to achieve better results. This approach involves generating instructions that are closely related to the source language. By providing more accurate and clear guidance for supervised fine-tuning of models, this technique enhances translation quality.
+
+![](images/dd82a0a7d282247a7d349f9fd5efa33054e6d5b87d850d43ffe79ca6a063b03c.jpg)  
+(b) Source-Language Consistent Instruction  
+Figure 3: Instruction Format
+
+To illustrate this concept, let's consider translations in English  $\Leftrightarrow$  Chinese and English  $\Leftrightarrow$  German. Traditional approaches typically employ a standardized English-Fixed instruction such as "Translate this sentence from the source language to the target language:". However, in Source-Language Consistent Instruction, the instruction varies based on the language pair. For English-to-Chinese translation, the instruction would be "把这句话从中文翻译成英文：" (Translate this sentence from Chinese to English). Similarly, for German-to-English translation, the instruction would be "Übersetzen Sie die folgenden Sätze vom Deutschen ins Englische:" (Translate the following sentences from German to English). By utilizing language-specific instructions, there is a semantic consistency established between the instruction and the source language, resulting in clearer and more accurate guidance. As shown in Figure 3.
+
+Supervised Instruction Fine-Tuning Supervised Instruction Fine-Tuning for machine translation tasks incorporates two pivotal aspects. Firstly, akin to the earlier phase of Continual Pre-training, we employ LoRA(Hu et al., 2021) to finely tune specific parameters of Language Learning Models (LLMs), thereby enhancing their efficiency. LoRA(Hu et al., 2021) plays a crucial role in preventing model overfitting and leads to notable performance improvements. With this approach, we judiciously fine-tune a subset of model parameters using low-rank updates, striking a delicate balance between model adaptation and computational effi
+
+ciency.
+
+Secondly, as emphasized in prior studies(Zhou et al., 2023; Maillard et al., 2023; Xu et al., 2023), LLMs exhibit benefits from a limited yet high-quality dataset. To ensure optimal data quality during the fine-tuning process, we leverage exceptional data sources. In line with previous research, we make use of meticulously curated human-written datasets derived from the WMT test data. These datasets undergo rigorous quality control measures, rendering them an ideal choice for fine-tuning purposes.
+
+# 4 Experiments
+
+# 4.1 Datasets and Evaluation Metrics
+
+The overall data statistics are shown in Table 5 of Appendix A. For Stage 2, we utilized the WMT bilingual training dataset consisting of English  $\Leftrightarrow$  German (En  $\Leftrightarrow$  De) and English  $\Leftrightarrow$  Chinese (En  $\Leftrightarrow$  Zh) sentence pairs. The En  $\Leftrightarrow$  De dataset comprised approximately 4.5 million pairs, while the En  $\Leftrightarrow$  Zh dataset contained around 25 million pairs. Due to the higher number of En  $\Leftrightarrow$  Zh pairs compared to En  $\Leftrightarrow$  De, we sampled 4.5 million En  $\Leftrightarrow$  Zh pairs for our experiments. Overall, the combined dataset contained nearly 1 billion tokens.
+
+For Stage 3, we employed the newstest2017-2020 dataset for both  $\mathrm{En} \Leftrightarrow \mathrm{Zh}$  and  $\mathrm{En} \Leftrightarrow \mathrm{De}$  translation tasks. This dataset included a total of 37.6 thousand sentence pairs for each language direction. To ensure consistency across the source language and target language, we organize these sentence pairs into Source-Language Consistent Instructions.
+
+We additionally incorporate the test sets from the WMT22 competition, which are carefully curated to include more recent content from diverse domains such as news, social media, e-commerce, and conversations. The test sets for the  $\mathrm{De} \Rightarrow \mathrm{En}$ ,  $\mathrm{En} \Rightarrow \mathrm{De}$ ,  $\mathrm{Zh} \Rightarrow \mathrm{En}$ , and  $\mathrm{En} \Rightarrow \mathrm{Zh}$  tasks consist of 1984, 2037, 1875, and 2037 samples, respectively.
+
+For automatic evaluation, we utilize Sacre-BLEU, which implements BLEU(Papineni et al., 2002), and COMET(Rei et al., 2020) from Unbabel/wmt22-comet-da. SacreBLEU calculates similarity based on n-gram matching, while COMET leverages cross-lingual pretrained models for evaluation.
+
+# 4.2 Setup
+
+We conducted our experiments using HuggingFace Transformers with open-source LLMs from the LLaMA(Touvron et al., 2023) family. Specifically, we utilized LLaMA2-7b with matched parameters as our foundation model. Additionally, we included LLaMA2-13b to explore the impact of different model sizes.
+
+Due to computational constraints, we did not reproduce the foundational experiments from Stage 1. After Stage 1, we selected Chinese-LLaMA2(Cui et al., 2023) as our new foundation model. Chinese-LLaMA2 is an extended and optimized version of Llama-2, specifically tailored for Chinese language understanding and instruction comprehension. It incorporates a larger Chinese vocabulary and undergoes incremental pretraining on a large-scale Chinese dataset, which further enhances its semantic understanding capabilities.
+
+For Stage 2, Continual Pre-training, and Stage 3, Supervised Fine-Tuning, we referred to the hyperparameters employed in the Chinese-LLaMA2 project. During Stage 2, we trained the model for 1 epoch, and for Stage 3, we extended the training to 3 epochs.
+
+Our experiments were conducted on 8 Nvidia GPUs with 64GB of memory each, utilizing DeepSpeed(Rasley et al., 2020) ZeRO 2 for model parallelization.
+
+# 4.3 Baselines
+
+We evaluate our method against two baseline categories, consistent with previous studies. Firstly, we compare our approach to prior studies that share our goal of leveraging LLMs for translation. Secondly, we benchmark against the current state-of-the-art (SoTA) translation models. It's important to note that this comparison may not be entirely fair due to disparities in training data and model architectures. For example, there is a significant contrast between the 175B GPT-3.5 model and our 7B model. Nevertheless, by utilizing the same test set, we can gain insights into our model's current performance.
+
+In the category of prior similar work, we compare our model to the following approaches: BigTranslate(Yang et al., 2023), which extends LLaMA-1-13B to cover over 100 translation directions; TIM(Zeng et al., 2023), which leverages correct and incorrect examples to aid LLMs in learning translation; ParroT(Jiao et al., 2023), through
+
+<table><tr><td rowspan="2">Models</td><td colspan="2">De→En</td><td colspan="2">En→De</td><td colspan="2">Zh→En</td><td colspan="2">En→Zh</td></tr><tr><td>BLEU</td><td>COMET</td><td>BLEU</td><td>COMET</td><td>BLEU</td><td>COMET</td><td>BLEU</td><td>COMET</td></tr><tr><td colspan="9">SoTA models</td></tr><tr><td>NLLB-54B(Team et al., 2022)</td><td>26.89</td><td>78.94</td><td>34.50</td><td>86.45</td><td>16.56</td><td>70.70</td><td>27.38</td><td>78.91</td></tr><tr><td>NLLB-54B Fine-tune</td><td>27.34</td><td>79.86</td><td>35.07</td><td>86.95</td><td>17.26</td><td>71.35</td><td>27.89</td><td>80.13</td></tr><tr><td>GPT-3.5-D, zero-shot</td><td>30.90</td><td>84.79</td><td>31.80</td><td>85.61</td><td>25.00</td><td>81.60</td><td>38.30</td><td>85.76</td></tr><tr><td>GPT-3.5-T, zero-shot</td><td>33.10</td><td>85.50</td><td>34.40</td><td>87.00</td><td>26.60</td><td>82.90</td><td>44.90</td><td>87.00</td></tr><tr><td>GPT-4</td><td>33.87</td><td>85.62</td><td>35.38</td><td>87.44</td><td>27.20</td><td>82.79</td><td>43.98</td><td>87.49</td></tr><tr><td colspan="9">Prior Similar Studies</td></tr><tr><td>TIM-7B(Zeng et al., 2023)</td><td>27.91</td><td>82.80</td><td>25.59</td><td>82.56</td><td>19.33</td><td>75.46</td><td>19.33</td><td>75.46</td></tr><tr><td>Parrot-7B(Jiao et al., 2023)</td><td>29.80</td><td>83.00</td><td>26.10</td><td>81.60</td><td>20.20</td><td>75.90</td><td>30.30</td><td>80.30</td></tr><tr><td>SWIE-7B(Chen et al., 2023)</td><td>30.48</td><td>82.97</td><td>27.21</td><td>82.36</td><td>21.30</td><td>76.48</td><td>31.24</td><td>80.63</td></tr><tr><td>ALMA-7B(Xu et al., 2023)</td><td>29.56</td><td>83.95</td><td>30.31</td><td>85.59</td><td>23.64</td><td>79.78</td><td>36.48</td><td>85.05</td></tr><tr><td>Parrot-13B(Jiao et al., 2023)</td><td>31.10</td><td>83.60</td><td>28.10</td><td>82.60</td><td>21.70</td><td>76.70</td><td>31.70</td><td>81.00</td></tr><tr><td>BigTranslate-13B(Yang et al., 2023)</td><td>23.35</td><td>80.68</td><td>21.48</td><td>78.81</td><td>14.16</td><td>74.26</td><td>28.56</td><td>81.31</td></tr><tr><td>Bayling-13B(Zhang et al., 2023)</td><td>27.34</td><td>83.02</td><td>25.62</td><td>82.69</td><td>20.12</td><td>77.72</td><td>37.92</td><td>84.62</td></tr><tr><td>ALMA-13B(Xu et al., 2023)</td><td>31.14</td><td>84.56</td><td>31.47</td><td>85.62</td><td>25.46</td><td>80.21</td><td>39.84</td><td>85.96</td></tr><tr><td>Ours</td><td colspan="8">Our Recipe with Backbone Model: LLaMA2(Touvron et al., 2023)</td></tr><tr><td>7B Stage3</td><td>30.02</td><td>84.09</td><td>25.40</td><td>82.30</td><td>20.59</td><td>76.18</td><td>30.60</td><td>80.40</td></tr><tr><td>7B Stage1,3*</td><td>25.20</td><td>78.32</td><td>12.50</td><td>69.19</td><td>20.90</td><td>76.40</td><td>35.00</td><td>84.32</td></tr><tr><td>7B Stage2,3</td><td>31.14</td><td>84.70</td><td>30.50</td><td>85.62</td><td>21.97</td><td>78.45</td><td>39.00</td><td>85.79</td></tr><tr><td>7B Stage1,2,3*</td><td>30.10</td><td>83.96</td><td>29.90</td><td>83.86</td><td>22.20</td><td>79.88</td><td>41.10</td><td>86.37</td></tr><tr><td>13B Stage3</td><td>31.70</td><td>84.39</td><td>28.80</td><td>83.87</td><td>21.40</td><td>77.68</td><td>35.90</td><td>84.23</td></tr><tr><td>13B Stage1,3*</td><td>26.13</td><td>78.65</td><td>12.79</td><td>72.23</td><td>21.40</td><td>78.28</td><td>37.34</td><td>85.27</td></tr><tr><td>13B Stage2,3</td><td>32.24</td><td>85.17</td><td>32.53</td><td>86.14</td><td>22.57</td><td>79.05</td><td>40.40</td><td>85.98</td></tr><tr><td>13B Stage1,2,3*</td><td>30.21</td><td>84.26</td><td>30.41</td><td>84.72</td><td>23.10</td><td>80.53</td><td>42.30</td><td>86.65</td></tr></table>
+
+Table 1: The overall results. Note: * Due to computational constraints, we did not reproduce the foundational experiments from Stage 1, but instead directly utilized the Chinese-Llama2(Cui et al., 2023) that had undergone similar training. Since Chinese-Llama2(Cui et al., 2023) was only trained in Chinese during Stage 1, our main analysis about Stage 1 focuses on its performance in  $\mathrm{{Zh}} \Rightarrow  \mathrm{{En}}$  and  $\mathrm{{En}} \Rightarrow  \mathrm{{Zh}}$  translations.
+
+three types of instructions including translation instruction, contrastive instruction, and error-guided instruction, improves the translation performance of LLM after SFT; SWIE(Chen et al., 2023), which enhances LLMs in translation through instruction augmentation; BayLing(Zhang et al., 2023), which incorporates interactive translation instructions; and ALMA(Xu et al., 2023), a two-stage finetuning method that initially fine-tunes on monolingual data and subsequently on a small set of high-quality parallel data.
+
+In the SoTA models category, we consider the following: the NLLB-54B(Team et al., 2022) model, the largest and best translation model released in the NLLB family; the zero-shot performance of GPT3.5-text-davinci-003 (GPT-3.5-D) and GPT-3.5-turbo-0301 (GPT-3.5-T). Additionally, we present the zero-shot results for GPT-4. For a fair comparison, we also compared the results of fine-tuning NLLB-54B model with 37.6k data in Stage 3. To evaluate these baselines, we employ the same test data and evaluation metrics, reporting BLEU(Papineni et al., 2002) and COMET(Rei et al., 2020) scores as provided in their respective
+
+papers.
+
+# 5 Results and Analysis
+
+As shown in Table 1, overall, our results outperform all previous studies, NLLB-54B(Team et al., 2022), and GPT-3.5-D, except for a slight lag in  $\mathrm{Zh}\Rightarrow \mathrm{En}$ . Even our 7B model surpasses the results of other works. Particularly in the  $\mathrm{En}\Rightarrow \mathrm{Zh}$  direction, our BLEU score is approximately 2.5 higher than the previous state-of-the-art. These findings are a testament to the effectiveness of our approach.
+
+# 5.1 Assessing the Impact of Stage 1
+
+Just as mentioned earlier, we didn't specifically train Llama2 in Stage 1, but instead directly utilized the Chinese-Llama2(Cui et al., 2023) that had undergone similar training. Since Chinese-Llama2(Cui et al., 2023) was only trained in Chinese during Stage 1, our main analysis focuses on its performance in  $\mathrm{Zh}\Rightarrow \mathrm{En}$  and  $\mathrm{En}\Rightarrow \mathrm{Zh}$  translations.
+
+As shown in Table 1, our findings align with previous research conclusions that incremental train
+
+ing on monolingual data is beneficial. Furthermore, we discovered that this benefit primarily affects the target language in translation tasks. For example, we observed a significant improvement in the performance of the 7B model on the  $\mathrm{En} \Rightarrow \mathrm{Zh}$  test set, where the BLEU score increased from 30.60 to 35.00, a substantial improvement of 4.4 points. However, the improvement in the  $\mathrm{Zh} \Rightarrow \mathrm{En}$  direction was limited, indicating that the role of Stage 1 is to enhance generation rather than comprehension.
+
+Additionally, we found that performing incremental training on only one monolingual dataset had disastrous effects on translation tasks in other languages. For example, on the  $\mathrm{En} \Rightarrow \mathrm{De}$  test set, the BLEU score plummeted from 25.40 to 12.50. Therefore, for multilingual translation, it is crucial to conduct Stage 1 training on multiple languages.
+
+# 5.2 Measuring the Effectiveness of Stage 2
+
+As shown in Table 1, Llama2(Touvron et al., 2023) demonstrates improved quality across various test sets after Stage 2 training. An interesting observation, considering Llama2 as a large model primarily focused on English, is that the enhancement in English-Other translations is particularly noteworthy after Stage 2 Training. For instance, the 7B model exhibits remarkable improvements in  $\mathrm{En}\Rightarrow \mathrm{De}$ , with the BLEU score increasing from 25.40 to 30.50, and in  $\mathrm{En}\Rightarrow \mathrm{Zh}$ , where it rises from 30.60 to 39.00. The magnitude of these improvements is quite significant. Encouragingly, there are also improvements observed in translations from other languages to English.
+
+An even more intriguing finding is that, as mentioned before, since Chinese-Llama2(Cui et al., 2023) only underwent Stage 1 training on Chinese, its translation performance substantially deteriorates in the  $\mathrm{En} \Rightarrow \mathrm{De}$  direction. However, with the magical touch of Stage 2 training, these capabilities are miraculously restored. The 7B model, on  $\mathrm{En} \Rightarrow \mathrm{De}$ , rebounds from 12.50 to 29.90, approaching the results of the original Llama2(Touvron et al., 2023). These outcomes effectively affirm the effectiveness of Stage 2.
+
+After considering the overall process, we are interested in understanding the impact of Stage 2 only. As mentioned before, LLMs typically include two main types of models: Foundation Models and Chat Models. Evaluating Stage 2 essentially assesses the Foundation Model by using an n-shot evaluation, which includes both zero-shot and 5-shot evaluations. We have noticed that zero-shot
+
+evaluations can occur hallucinations. Hence, we are presenting the results of the 5-shot evaluation in Table 2.
+
+<table><tr><td>Models</td><td colspan="2">Zh→En</td><td colspan="2">En→Zh</td></tr><tr><td></td><td>BLEU</td><td>COMET</td><td>BLEU</td><td>COMET</td></tr><tr><td>Baseline</td><td>20.63</td><td>76.32</td><td>29.96</td><td>79.34</td></tr><tr><td>+ Stage2</td><td>21.64</td><td>78.07</td><td>38.62</td><td>85.30</td></tr></table>
+
+# 5.3 Analyzing the Outcomes of Stage 3
+
+To evaluate the effectiveness of Source-Language Consistent Instruction in Stage 3, we conducted a comparative experiment using English-Fixed Instruction. The results of the experiment are presented in Table 3. It is evident that in the  $\mathrm{En}\Rightarrow \mathrm{De}$  and  $\mathrm{De}\Rightarrow \mathrm{En}$  directions, the performance of these two types of instructions is quite similar. However, in the  $\mathrm{Zh}\Rightarrow \mathrm{En}$  and  $\mathrm{En}\Rightarrow \mathrm{Zh}$  directions, the use of Source-Language Consistent Instruction clearly outperforms.
+
+Table 2: Results of the five-shot results based on Llama2-7B(Touvron et al., 2023) model.  
+
+<table><tr><td>Models</td><td>De⇒En</td><td>En⇒De</td><td>Zh⇒En</td><td>En⇒Zh</td></tr><tr><td>Stage 3</td><td>30.02</td><td>25.40</td><td>20.59</td><td>30.60</td></tr><tr><td>w/o</td><td>30.40</td><td>25.20</td><td>18.39</td><td>28.30</td></tr><tr><td>Stage2,3</td><td>31.14</td><td>30.50</td><td>21.91</td><td>39.00</td></tr><tr><td>w/o</td><td>31.00</td><td>30.23</td><td>18.93</td><td>38.69</td></tr></table>
+
+Table 3: Results of the comparative experiments based on Llama2-7B(Touvron et al., 2023) model. [w/o] means using English-Fixed Instruction.
+
+We believe that the similarity between English and German, as they belong to the same language family, contributes to the lack of noticeable differences. However, when dealing with cross-language pairs, employing Source-Language Consistent Instruction further enhances the translation quality.
+
+# 5.4 Comparing the Difference with ALMA
+
+We have noticed that our work shares some similarities with ALMA(Xu et al., 2023) in terms of the process, involving Continual Pre-training followed by Supervised Fine-Tuning. However, there are notable differences between our approaches.
+
+ALMA suggests that the impact of bilingual data is reduced in the era of large models. In contrast, we firmly believe in the continued strength of bilingual data and its application in Continual Pre-training through Interlinear Text Format Documents. While ALMA acknowledges the effective
+
+ness of conducting Continual Pre-training on monolingual data, we have also validated this finding in our own work and reached the same conclusion. However, it is important to note that our approach primarily enhances the multilingual generation capability of the large model itself, rather than being specifically tailored to translation tasks. Furthermore, ALMA utilizes a significantly larger training dataset, ranging from 13B to 20B, compared to our own.
+
+# 6 Ablation Study: What if we directly employ a large quantity of translation data for SFT?
+
+Both Continual Pre-training and Supervised Fine-Tuning involve incremental training on the original model. However, if we skip Stage 2 training and directly utilize the translation data from Stage 2 as instruction data for SFT, i.e., conducting SFT directly with a substantial amount of translation data, will it yield consistent improvement?
+
+<table><tr><td>Data Size</td><td>De→En</td><td>En→De</td><td>Zh→En</td><td>En→Zh</td></tr><tr><td>37.6K</td><td>30.02</td><td>25.40</td><td>20.59</td><td>30.60</td></tr><tr><td>400K</td><td>30.20</td><td>25.60</td><td>18.49</td><td>31.74</td></tr><tr><td>4,000K</td><td>30.66</td><td>25.12</td><td>20.77</td><td>32.22</td></tr></table>
+
+Table 4: Results of the ablation experiments based on Llama2-7B(Touvron et al., 2023) model under different Stage 3 data size.
+
+To address this question, we conducted an ablation experiment. Our Stage 3 training data consisted of  $37.6\mathrm{k}$  samples. Randomly selecting and merging some data from the Stage 2 training data with the Stage 3 training data, we created three sets:  $37.6\mathrm{K},400\mathrm{K}$ , and  $4,000\mathrm{K}$ . The experimental results are presented in Table 4.
+
+We found that augmenting the training data in Stage 3 slightly improved translation quality for certain test sets. This indicates that a small amount of high-quality data is sufficient for the SFT stage.
+
+Now, our focus is solely on the translation task. However, if we were conducting multi-task SFT, it is unlikely that other tasks would have as extensive data as machine translation. Therefore, using a large amount of translation data during SFT would result in the problem of imbalanced data distribution with other tasks. Hence, the optimal approach would still be to utilize this substantial data during the earlier stage of Continual Pre-training.
+
+# 7 Conclusions
+
+In this study, we have introduced a novel paradigm for enhancing the translation capabilities of large language models in machine translation tasks. Our three-stage approach, including Secondary Pretraining using Extensive Monolingual Data, Continual Pre-training with Interlinear Text Format Documents, and Leveraging Source-Language Consistent Instruction for Supervised Fine-Tuning, addresses the limitations of previous strategies and offers notable improvements in translation quality. We emphasize the significance of pre-training stages in enhancing LLMs' cross-lingual alignment abilities and the effectiveness of using a smaller but high-quality set of bilingual data during supervised fine-tuning. Notably, Stage2, which involves Continual Pre-training with Interlinear Text Format Documents, stands out as a highly efficient method, requiring minimal training data. Furthermore, aligning the instructional setting with the source language during supervised fine-tuning, as observed in Stage3, yields positive effects. The findings from this paper contribute to advancing the field of machine translation and offer valuable insights for optimizing the translation capabilities of large language models. Future research can explore additional language pairs, alternative data augmentation techniques, and different pre-training strategies to further refine our proposed paradigm.
+
+# 8 Limitations
+
+Despite notable contributions, this study has certain limitations. Firstly, the proposed method exhibits slightly reduced performance in the  $\mathrm{Zh} \Rightarrow \mathrm{En}$  translation direction, necessitating further analysis and improvements. Secondly, the presence of illusionary translations within large models was observed but not extensively explored. Future research should delve deeper into this phenomenon. Lastly, while the paper primarily focuses on SFT for machine translation, opportunities exist to explore SFT techniques in diverse contexts such as style translation and colloquial translation. Addressing these limitations would further enhance the effectiveness and applicability of the proposed methods.
+
+# References
+
+Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind
+
+Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. CoRR, abs/2005.14165.  
+Yijie Chen, Yijin Liu, Fandong Meng, Yufeng Chen, Jinan Xu, and Jie Zhou. 2023. Improving translation faithfulness of large language models via augmenting instructions.  
+Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. 2023. Palm: Scaling language modeling with pathways. J. Mach. Learn. Res., 24:240:1-240:113.  
+Yiming Cui, Ziqing Yang, and Xin Yao. 2023. Efficient and effective text encoding for chinese llama and alpaca. arXiv preprint arXiv:2304.08177.  
+Marjan Ghazvininejad, Omer Levy, Yinhan Liu, and Luke Zettlemoyer. 2019. Mask-predict: Parallel decoding of conditional masked language models. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, November 3-7, 2019, pages 6111-6120. Association for Computational Linguistics.  
+Jiatao Gu, James Bradbury, Caiming Xiong, Victor O. K. Li, and Richard Socher. 2018. Non-autoregressive neural machine translation. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net.  
+Jiaxin Guo, Minghan Wang, Daimeng Wei, Hengchao Shang, Yuxia Wang, Zongyao Li, Zhengzhe Yu, Zhanglin Wu, Yimeng Chen, Chang Su, Min Zhang,
+
+Lizhi Lei, Shimin Tao, and Hao Yang. 2021. Self-distillation mixup training for non-autoregressive neural machine translation. CoRR, abs/2112.11640.  
+Amr Hendy, Mohamed Abdelrehim, Amr Sharaf, Vikas Raunak, Mohamed Gabr, Hitokazu Matsushita, Young Jin Kim, Mohamed Affify, and Hany Hassan Awadalla. 2023. How good are GPT models at machine translation? A comprehensive evaluation. CoRR, abs/2302.09210.  
+Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models.  
+Wenxiang Jiao, Jen tse Huang, Wenxuan Wang, Zhiwei He, Tian Liang, Xing Wang, Shuming Shi, and Zhaopeng Tu. 2023. Parrot: Translating during chat using large language models tuned with human translation and feedback.  
+Jean Maillard, Cynthia Gao, Elahe Kalbassi, Kaushik Ram Sadagopan, Vedanuj Goswami, Philipp Koehn, Angela Fan, and Francisco Guzmán. 2023. Small data, big impact: Leveraging minimal data for effective machine translation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2023, Toronto, Canada, July 9-14, 2023, pages 2740-2756. Association for Computational Linguistics.  
+Kishore Papineni, Salim Roukos, Todd Ward, and Wei Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, July 6-12, 2002, Philadelphia, PA, USA, pages 311-318. ACL.  
+Jeff Rasley, Samyam Rajbhandari, Olatunjri Ruwase, and Yuxiong He. 2020. Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In KDD '20: The 26th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Virtual Event, CA, USA, August 23-27, 2020, pages 3505-3506. ACM.  
+Ricardo Rei, Craig Stewart, Ana C. Farinha, and Alon Lavie. 2020. COMET: A neural framework for MT evaluation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020, pages 2685-2702. Association for Computational Linguistics.  
+Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Improving neural machine translation models with monolingual data.  
+NLLB Team, Marta R. Costa-jussa, James Cross, Onur Celebi, Maha Elbayad, Kenneth Heafield, Kevin Hefernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, Anna Sun, Skyler Wang, Guillaume Wenzek, Al Youngblood, Bapi Akula, Loic Barrault, Gabriel Mejia Gonzalez, Prangthip Hansanti,
+
+John Hoffman, Semarley Jarrett, Kaushik Ram Sadagopan, Dirk Rowe, Shannon Spruit, Chau Tran, Pierre Andrews, Necip Fazil Ayan, Shruti Bhosale, Sergey Edunov, Angela Fan, Cynthia Gao, Vedanuj Goswami, Francisco Guzmán, Philipp Koehn, Alexandre Mourachko, Christophe Ropers, Safiyyah Saleem, Holger Schwenk, and Jeff Wang. 2022. No language left behind: Scaling human-centered machine translation.
+
+Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. 2023. Llama 2: Open foundation and finetuned chat models.
+
+Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5998-6008.
+
+Minghan Wang, Jiaxin Guo, Yuxia Wang, Yimeng Chen, Chang Su, Daimeng Wei, Min Zhang, Shimin Tao, and Hao Yang. 2021. HI-CMLM: improve CMLM with hybrid decoder input. In Proceedings of the 14th International Conference on Natural Language Generation, INLG 2021, Aberdeen, Scotland, UK, 20-24 September, 2021, pages 167-171. Association for Computational Linguistics.
+
+Daimeng Wei, Zhanglin Wu, Hengchao Shang, Zongyao Li, Minghan Wang, Jiaxin Guo, Xiaoyu Chen, Zhengzhe Yu, and Hao Yang. 2023. Text style transfer back-translation.
+
+Haoran Xu, Young Jin Kim, Amr Sharaf, and Hany Hassan Awadalla. 2023. A paradigm shift in machine translation: Boosting translation performance of large language models.
+
+Wen Yang, Chong Li, Jiajun Zhang, and Chengqing Zong. 2023. Bigtranslate: Augmenting large language models with multilingual translation capability over 100 languages.
+
+Zhengzhe Yu, Jiaxin Guo, Minghan Wang, Daimeng Wei, Hengchao Shang, Zongyao Li, Zhanglin Wu, Yuxia Wang, Yimeng Chen, Chang Su, Min Zhang, Lizhi Lei, Shimin Tao, and Hao Yang. 2021. Jointtraining on symbiosis networks for deep nerval machine translation models. CoRR, abs/2112.11642.
+
+Jiali Zeng, Fandong Meng, Yongjing Yin, and Jie Zhou. 2023. Tim: Teaching large language models to translate with comparison.
+
+Shaolei Zhang, Qingkai Fang, Zhuocheng Zhang, Zhengrui Ma, Yan Zhou, Langlin Huang, Mengyu Bu, Shangtong Gui, Yunji Chen, Xilin Chen, and Yang Feng. 2023. Bayling: Bridging cross-lingual alignment and instruction following through interactive translation for large language models.
+
+Chunting Zhou, Pengfei Liu, Puxin Xu, Srini Iyer, Jiao Sun, Yuning Mao, Xuezhe Ma, Avia Efrat, Ping Yu, Lili Yu, Susan Zhang, Gargi Ghosh, Mike Lewis, Luke Zettlemoyer, and Omer Levy. 2023. LIMA: less is more for alignment. CoRR, abs/2305.11206.
+
+# A Appendix A: Data Statistics
+
+Table 5 displays the comprehensive data statistics. For Stage 2, we utilized the WMT bilingual training dataset that includes English  $\Leftrightarrow$  German (En  $\Leftrightarrow$  De) and English  $\Leftrightarrow$  Chinese (En  $\Leftrightarrow$  Zh) sentence pairs. The En  $\Leftrightarrow$  De dataset comprised approximately 4.5 million pairs, while for the En  $\Leftrightarrow$  Zh dataset, we randomly sampled an equivalent number of pairs from the total 25 million pairs. In total, the combined dataset contained close to 1B tokens.
+
+Moving on to Stage 3, we utilized the newstest2017-2020 dataset for both the  $\mathrm{En} \Leftrightarrow \mathrm{De}$  and  $\mathrm{En} \Leftrightarrow \mathrm{Zh}$  translation tasks. This dataset comprised 37.6 thousand sentence pairs for each language direction. To maintain coherence between the source and target languages, we categorized these sentence pairs into Source-Language Consistent Instructions.
+
+# B Appendix B: Stage 2 Training Data Sample
+
+Figure 4 displays samples of the Stage 2 training data.
+
+<table><tr><td>Stage</td><td>Data</td><td>Description</td></tr><tr><td>Stage 1</td><td>120G text</td><td>120G Chinese text described in Chinese-LLaMA2(Cui et al., 2023)</td></tr><tr><td>Stage 2</td><td>1B tokens</td><td>1B tokens including four directions: En↔De and En↔Zh. Each direction includes 4.5 million pairs.</td></tr><tr><td>Stage 3</td><td>37.6k pairs</td><td>37.6k pairs combined wmt newstest2017-2020 testset with all four directions.</td></tr></table>
+
+Table 5: The overall data statistics.
+
+<s> But buses are faster and more frequent. \n Aber Buses sind Schneller und haufter. </s> <s> 因此, 公布的历史成本不反映非消耗性财产充分而准确的估价; \n Accordingly, the historical costs disclosed did not reflect the full and accurate valuation of non-expendable property; </s> Wir haben heute dem Bericht Ferber zugestimmt, obwohl er eine Grundlage für den Kauf der Parlamentsgebäude in Straßburg bildet. Unsere Zustimmung stellt aber keine Vorfestlegung auf den Sitz des Europäischen Parlaments dar. \n We voted in favour of the Ferber report today, even though it forms a basis for the purchase of the Parliament building in Strasbourg. However, our approval does not constitute a prior decision regarding the seat of the European Parliament. </s> <s> By introducing greater flexibility in how the Fund is used and by reducing the number of redundancies from 1000 to 500, it will become an ever more effective instrument for helping to tackle the effects of the economic down-turn. \n Durch die Einführung von mehr Flexibilität in der Nutzung der Fonds und durch die Reduzierung der Zahl der Entlassungen von 1000 auf 500 wird er ein noch effizienteres Instrument zur Bekämpfung der Folgen des Wirtschaftssabschwungs werden. </s> <s> During the period from 1 January to 30 June 2010, contributions totalling $19,299.39 were received from Angola ($10,000) and Congo ($9,299.39). \n 2010年1月1日至6月30日期间,从安哥拉(10000美元)和刚果(9299.39美元)共收到19299.39美元捐款。 </s> <s> ...
+
+Figure 4: Stage 2 Training Data Sample

2024/A Novel Paradigm Boosting Translation Capabilities of Large Language Models/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ff35415b76f026d3494843a198477c622fcdc7e518ddc9ac1fd87befceced6ed
+size 402846

2024/A Novel Two-step Fine-tuning Framework for Transfer Learning in Low-Resource Neural Machine Translation/07401a9f-9a22-4566-a6eb-41c799e41ff7_content_list.json

@@ -0,0 +1,1585 @@
+[
+    {
+        "type": "text",
+        "text": "A Novel Two-step Fine-tuning Framework for Transfer Learning in Low-Resource Neural Machine Translation",
+        "text_level": 1,
+        "bbox": [
+            149,
+            89,
+            847,
+            130
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Yuan Gao, Feng Hou*, Ruili Wang",
+        "bbox": [
+            349,
+            155,
+            653,
+            172
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "School of Mathematical and Computational Science, Massey University, New Zealand",
+        "bbox": [
+            149,
+            174,
+            850,
+            192
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{y.gao, f.hou, ruili.wang}@massey.ac.nz",
+        "bbox": [
+            265,
+            195,
+            737,
+            212
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            261,
+            339,
+            275
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Existing transfer learning methods for neural machine translation typically use a well-trained translation model (i.e., a parent model) of a high-resource language pair to directly initialize a translation model (i.e., a child model) of a low-resource language pair, and the child model is then fine-tuned with corresponding datasets. In this paper, we propose a novel two-step fine-tuning (TSFT) framework for transfer learning in low-resource neural machine translation. In the first step, we adjust the parameters of the parent model to fit the child language by using the child source data. In the second step, we transfer the adjusted parameters to the child model and fine-tune it with a proposed distillation loss for efficient optimization. Our experimental results on five low-resource translations demonstrate that our framework yields significant improvements over various strong transfer learning baselines. Further analysis demonstrated the effectiveness of different components in our framework.",
+        "bbox": [
+            141,
+            282,
+            460,
+            595
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            602,
+            260,
+            618
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Neural machine translation (NMT) has achieved superior performance in terms of both fluency and adequacy for high-resource languages (Vaswani et al., 2017; Zhou and Keung, 2020; Cai et al., 2021; Guo et al., 2022). With the introduction of the attention mechanism (Yin et al., 2021; Petrick et al., 2022), NMT has been proven to be efficient and powerful in modeling long-distance dependencies. However, the performance of NMT systems deteriorates dramatically when insufficient parallel data are available for training (Sakaguchi et al., 2017; Michel and Neubig, 2018; Aharoni et al., 2019; Goyal et al., 2022). The scarcity of parallel corpora intensely limits the performance of an NMT system on low-resource languages.",
+        "bbox": [
+            112,
+            627,
+            489,
+            868
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Transfer learning is a learning paradigm for addressing the data scarcity problem (Zoph et al.,",
+        "bbox": [
+            112,
+            869,
+            489,
+            902
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/dd040f27afdcbd4a07e347deab3de7619eadc964a52b708a4c36518fb20fb5c1.jpg",
+        "image_caption": [
+            "Figure 1: Comparison between vanilla transfer learning framework (a) and TSFT (b). Our proposed TSFT incorporates an intermediate model to pre-fine-tune the parent parameters to fit the child data."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            539,
+            263,
+            863,
+            456
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "2016; Nguyen and Chiang, 2017; Li et al., 2022). For NMT, transfer learning aims to transfer the knowledge from a well-trained high-resource translation model (i.e., a parent model, e.g., English  $\\rightarrow$  German) to a low-resource translation model (i.e., a child model, e.g., English  $\\rightarrow$  the Māori language). Prior transfer learning methods in NMT (Zoph et al., 2016; Chu et al., 2017) primarily achieve knowledge transfer by initializing the parameters of the child model with the parent model and fine-tuning the child model on the corresponding data. Such direct transfer of knowledge raises a vocabulary mismatch problem (Lakew et al., 2018; Lin et al., 2019; Kocmi and Bojar, 2020), and results in unsatisfied results for low-resource translations. Some methods have been proposed to alleviate the vocabulary mismatch problem, such as constructing joint dictionaries or employing a crosslingual token mapping technique (Passban et al., 2017; Kocmi and Bojar, 2018; Kim et al., 2019a). Additionally, Aji et al. (2020) proposed a token matching method that simply duplicates the embed",
+        "bbox": [
+            507,
+            567,
+            885,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Corresponding author.",
+        "bbox": [
+            141,
+            906,
+            289,
+            920
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "3214",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: NAACL 2024, pages 3214-3224 June 16-21, 2024 ©2024 Association for Computational Linguistics",
+        "bbox": [
+            218,
+            945,
+            778,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "dings of overlapping tokens from the parent model to the child model.",
+        "bbox": [
+            112,
+            84,
+            485,
+            115
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Recently, based on the work of Aji et al. (2020), Li et al. (2022) proposed ConsistTL that uses the predictions of the parent model to continuously provide soft targets during the fine-tuning of the child model. However, given the differences between the source inputs of the parent and the child translation tasks, the parent model is not an optimal starting point for the single-step fine-tuning of the child model using limited parallel child data. Therefore, it is necessary to pre-fine-tune the parent model to fit the child language before initializing the child model with it.",
+        "bbox": [
+            112,
+            117,
+            487,
+            309
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Building upon this insight, we propose a simple yet effective transfer learning framework, named Two-Step Fine-Tuning (TSFT), for low-resource NMT. As shown in Figure 1, we introduce an intermediate (child) model initialized with the parent model to adjust the parent parameters to fit the child language. TSFT involves two fine-tuning steps. In the first step, we feed child source sentences (i.e., monolingual data) and meaning-matched sentences in the parent source language into the intermediate and the parent models, respectively. Then, the intermediate model is fine-tuned with the objective of aligning probability distributions from the parent and intermediate models, aiming to adjust the parameters transferred from the parent model to perform well with child source sentences. Additionally, we propose a regularization-based strategy that can improve the translation performance of the intermediate model and benefit the child model. Note that we apply the token matching method to alleviate the vocabulary mismatch problem in the first step. In the second step, we transfer the adjusted parameters from the intermediate model to the child model and fine-tune the entire child model on the pertinent parallel data, employing both a cross-entropy loss and a proposed distillation loss. Extensive experiments on five low-resource translations show that TSFT surpasses the strongest baseline method with up to 1.2 SacreBLEU points. The ablation study demonstrates the effectiveness of different components within TSFT.",
+        "bbox": [
+            112,
+            311,
+            489,
+            809
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our contributions can be summarized as follows:",
+        "bbox": [
+            131,
+            810,
+            487,
+            826
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "- We propose a novel two-step fine-tuning framework for low-resource NMT, which introduces an intermediate (child) model to fit parent parameters for the data of child languages before initializing the child model with",
+        "bbox": [
+            134,
+            841,
+            489,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "the parent model.",
+        "bbox": [
+            544,
+            84,
+            678,
+            99
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We propose a regularization-based strategy for fine-tuning the intermediate model and a distillation loss for fine-tuning the child model.",
+            "- We validate our method by extensive experiments on various low-resource translations and achieve improved performance compared to various transfer learning methods."
+        ],
+        "bbox": [
+            531,
+            112,
+            884,
+            237
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related work",
+        "text_level": 1,
+        "bbox": [
+            507,
+            250,
+            660,
+            265
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Existing studies have demonstrated the success of transfer learning for low-resource NMT (Lin et al., 2019; Imankulova et al., 2019; Ji et al., 2020; Eronen et al., 2023). Zoph et al. (2016) first introduced transfer learning into the field of NMT and proposed a parent-child framework, where parameters from a pre-trained parent model are directly transferred to a new child model with a shared target language. Subsequent research largely builds upon the parent-child framework and tends to leverage highly related parent language to perform transfer learning (Passban et al., 2017; Setiawan et al., 2018). However, the languages closely related to low-resource languages are also low-resourced (Nguyen and Chiang, 2017; Xia et al., 2019) and offer only modest performance improvements. Thus, researchers focused on identifying the critical factors for the effectiveness of the parent language. Experimental results from (Lin et al., 2019; Aji et al., 2020) emphasized that linguistic or geographical distance does not appear as important as the size of the parent data (Lin et al., 2019; Aji et al., 2020). This insight expands the range of parent languages available for transfer learning, and alleviates the limitations of highly related parent languages. Consequently, later researchers shifted their attention to parent languages with low relatedness but high-resourced. However, this exacerbates the vocabulary mismatch problem, posing a new challenge to transfer learning.",
+        "bbox": [
+            507,
+            275,
+            884,
+            759
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "One solution to the vocabulary mismatch problem is to build a joint dictionary before training a parent model (Kocmi and Bojar, 2018; Kim et al., 2019b). However, this restricts the applicability of a pre-trained parent model to a specific child model only. To overcome this limitation, Kim et al. (2019a) proposed pre-training a language-agnostic cross-lingual word embedding independently from the parent model. Concurrently, token matching methods also show their effectiveness in transfer",
+        "bbox": [
+            507,
+            760,
+            885,
+            920
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "3215",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "learning without requiring additional training efforts (Aji et al., 2020; Kocmi and Bojar, 2020). Some other methods introduce highly related intermediate languages to gradually narrow the vocabulary disparity (Luo et al., 2019; Maimaiti et al., 2019). These methods take advantage of both large-scale data sources and syntactic similarity in the intermediate language.",
+        "bbox": [
+            112,
+            84,
+            489,
+            212
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Recently, Li et al. (2022) incorporated the idea of consistency learning into transfer learning based on the work of Aji et al. (2020) and proposed a novel transfer learning method called ConsistTL. This method enables the child model to utilize the parent model during fine-tuning. Subsequently, Liu et al. (2023) proposed kNN-TL, which extends ConsistTL by integrating a k-nearest neighbor (kNN) module, allowing the child model to utilize the parent model during inference. While our method also builds on ConsistTL, we focus on enhancing the child model's performance during fine-tuning. Thus, our work is orthogonal to kNN-TL.",
+        "bbox": [
+            115,
+            214,
+            489,
+            437
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Method",
+        "text_level": 1,
+        "bbox": [
+            112,
+            450,
+            216,
+            464
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this section, we begin by providing an overview of the basic concepts behind transfer learning and then present our transfer learning framework, TSFT, in detail.",
+        "bbox": [
+            112,
+            475,
+            487,
+            539
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Transfer Learning Primary",
+        "text_level": 1,
+        "bbox": [
+            112,
+            550,
+            376,
+            565
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Given a source sentence  $x = \\{x_{1},\\ldots ,x_{I}\\}$ , the objective of an NMT model is to translate it to a new sentence  $y = \\{y_{1},\\dots ,y_{J}\\}$  in a target language, where the source sentence and target sentence have lengths  $I$  and  $J$ , respectively. A typical NMT model is composed of an encoder and a decoder. The encoder is designed to extract high-level semantic information from the source sentences and represent them as hidden states  $H_{e}$ . The decoder generates the output probability  $P(y_{i}|H_{e},y_{< i})$  of the next target token  $y_{i}$ . An NMT model is trained on a parallel corpus by minimizing the cross-entropy (CE) loss between the predicted sentence and the ground-truth translation as follows:",
+        "bbox": [
+            112,
+            571,
+            489,
+            810
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nL _ {c e} = - \\sum_ {i = 1} ^ {J} \\log P \\left(y _ {i} \\mid y _ {<   i}, x, \\theta\\right), \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            176,
+            824,
+            487,
+            866
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\theta$  is the parameters of the entire NMT model.",
+        "bbox": [
+            112,
+            873,
+            487,
+            888
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Transfer learning has been widely used when only limited training datasets are available for the",
+        "bbox": [
+            112,
+            889,
+            487,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "problem at hand. It transfers the knowledge acquired from large-scale data to enhance the model performance under low-resource conditions. Transfer learning typically follows a parent-child framework (Zoph et al., 2016), where it involves reusing the parameters  $\\theta_{p}$  from a pre-trained parent model to initialize part or all parameters of a child model. In the field of NMT, the parent model  $\\mathcal{M}_p$  is initially trained on a high-resourced parallel dataset  $D_{p} = \\{X_{p},Y_{p}\\}$ , while there is only a limited-sized dataset  $D_{c} = \\{X_{c},Y_{c}\\}$  available to the child model  $\\mathcal{M}_c$ . After the initialization step, the child model can be fine-tuned on  $D_{c}$ , which is also optimized through the minimization of the CE loss.",
+        "bbox": [
+            507,
+            84,
+            884,
+            309
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2 Two-step Fine-tuning",
+        "text_level": 1,
+        "bbox": [
+            507,
+            319,
+            724,
+            336
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "For NMT, an ideal transfer learning framework should enable the parent model to exert its complete capabilities on the child task. However, owing to the disparities between the parent and child languages, the current one-step fine-tuning transfer learning framework struggles to adjust the parameters of the parent model to fit the child source language under the constraints of limited child data.",
+        "bbox": [
+            507,
+            341,
+            882,
+            469
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The idea of TSFT is simple: before initializing the child model with the parent model, we first adjust the parameters of the parent model to enhance its congruity with the child source language. In this work, we propose to introduce an intermediate model, denoted as  $\\mathcal{M}_a$ , to make the parameters of the parent model fit for the child data. Specifically, we initialize the intermediate model with the parent model and pre-fine-tune it by using the source side sentences of the child data, then fine-tune the child model with both the source and target child training data. Therefore, we design TSFT as a two-step framework, as shown in Figure 2.",
+        "bbox": [
+            507,
+            470,
+            882,
+            696
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Step 1: Intermediate Fine-tuning After initializing the intermediate model with a well-trained parent model, we aim to equip the intermediate model with the ability to utilize child source sentences as input for target language generation. Since the intermediate model and the parent model share the same target language, it is crucial to retain the generation ability of the parent model. Therefore, we input the source-side sentences of the child data to the intermediate model and the parent model and utilize the predicted distribution of the parent model as the soft label for fine-tuning.",
+        "bbox": [
+            507,
+            712,
+            884,
+            904
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "However, it is infeasible to directly input child",
+        "bbox": [
+            526,
+            904,
+            880,
+            920
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "3216",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/26b1189479dd93c41352c7e01867b76afb5ca5c53a401e01682bf2ac98306d8f.jpg",
+        "image_caption": [
+            "Figure 2: Our proposed transfer learning framework TSFT for low-resource NMT. In Step 1, the loss function  $L_{inter}$  is used to optimize the intermediate model. In Step 2, the child model is optimized by  $L_{child}$ . The blue icy blocks are initialized with the parent model and frozen. The input German sentences are produced through back-translation."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            82,
+            884,
+            272
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "source sentences into the parent model, given that the parent and child models have different source languages. Thus, we need a meaning-matched sentence for each child source sentence in the parent source language. In the context of low-resource translations, parallel data for non-English-centric is often limited in size or entirely absent, making it difficult to meet the requirements for intermediate fine-tuning. Therefore, we adopt the method of Li et al. (2022) to generate pseudo parent data  $D_{p*} = \\{X_{p*}, Y_c\\}$  by using a reversed parent model, where each  $x_{p*} \\in X_{p*}$  is aligned with  $y_c \\in Y_c$ . Although such a method requires training a reverse parent model, it effectively generates meaning-matched input sentences for the parent model. In addition, we use the following loss function to optimize the intermediate model:",
+        "bbox": [
+            112,
+            363,
+            489,
+            640
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nL _ {i n t e r} = \\sum_ {i = 1} ^ {J} F _ {d} [ P _ {i n t e r} (y _ {i}), P _ {p a r e n t} (y _ {i}) ], \\quad (2)\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            137,
+            659,
+            487,
+            703
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $F_{d}$  is a distribution measurement method, in this work, we choose Jensen-Shannon (JS) divergence (Lin, 1991; Wen et al., 2023) as our  $F_{d}$ . Our preliminary experiments find that JS divergence outperforms using Kullback-Leibler (KL) divergence when taking  $P_{inter}(y_i)$  as the first item and  $P_{parent}(y_i)$  as the second one.  $P_{*}(y_i)$  represents the prediction distributions of translation models at time step  $i$ , which is conditioned on the input sentence and the previous tokens:",
+        "bbox": [
+            112,
+            712,
+            489,
+            873
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nP _ {*} (y _ {i}) = P _ {*} (y _ {i} | x, y _ {<   i}). \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            208,
+            903,
+            487,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Before fine-tuning the intermediate model, we first apply the token matching method (Aji et al., 2020) that duplicates the embeddings of overlapping tokens from the parent and child vocabularies to alleviate the vocabulary mismatch problem.",
+        "bbox": [
+            507,
+            363,
+            884,
+            461
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Step 2: Child Fine-tuning In the second step, we employ the target-side sentences from the child training data as labels to fine-tune the child model with CE loss, following the general process of transfer learning. Since the encoder of the intermediate model has fine-tuned with the child source sentences, we argue that it encompasses valuable information that can facilitate the child model. Therefore, we extract the encoder outputs,  $P_{*}^{e}(\\cdot)$ , from both the intermediate and child models and incorporate a distillation loss  $L_{dist}$  as an extra objective to optimize the child model by minimizing the KL divergence between two output representations:",
+        "bbox": [
+            507,
+            476,
+            885,
+            686
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nL _ {d i s t} = - \\sum_ {i = 1} ^ {I} P _ {i n t e r} ^ {e} (x _ {i}) \\cdot l o g P _ {c h i l d} ^ {e} (x _ {i}), \\quad (4)\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            532,
+            709,
+            882,
+            752
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} P _ {*} ^ {e} (x _ {i}) = P _ {*} ^ {e} (x _ {i} | x, \\tau) \\\\ = \\frac {\\exp \\left(z _ {i} / \\tau\\right)}{\\sum_ {j \\in V} \\exp \\left(z _ {j} / \\tau\\right)}, \\tag {5} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            584,
+            775,
+            880,
+            832
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $I$  denotes the sentence length of a child source sentence,  $z$  denotes the logits output of encoders before log softmax is computed,  $V$  represents the vocabulary, and  $\\tau$  is a temperate factor used to smooth the prediction distributions. As we",
+        "bbox": [
+            507,
+            841,
+            884,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "3217",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "only reuse the output of encoders, the process of encoder distillation does not add any extra parameters to models. The overall loss is obtained by a weighted sum of  $L_{ce}$  and  $L_{dist}$ :",
+        "bbox": [
+            112,
+            84,
+            489,
+            149
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nL _ {\\text {c h i l d}} = L _ {c e} + \\lambda L _ {\\text {d i s t}}, \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            211,
+            168,
+            487,
+            185
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\lambda$  is a balancing hyper-parameter.",
+        "bbox": [
+            112,
+            196,
+            413,
+            212
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Partial Decoder Freeze Regularization-based methods are widely used to alleviate the catastrophic forgetting issue (Kirkpatrick et al., 2017; Gu and Feng, 2020; Gu et al., 2021). While updating all parameters typically yields good results on a new domain, the data distribution difference between the old and new domains can engender the issue of catastrophic forgetting, causing the fine-tuned model to abandon linguistic knowledge learned from previous dataset (Thompson et al., 2019; Bérard, 2021). In this work, we are interested in introducing the regularization-based technique during Step 1 to preserve the predictive capabilities of the parent model. We propose a Partial Decoder Freeze (PDF) strategy to freeze the parameters of the last  $l$  decoder layers of the intermediate model and only update the rest parameters. For the selection of parameters  $l$ , we conducted empirical experiments in Section 5.1.",
+        "bbox": [
+            112,
+            229,
+            489,
+            533
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4 Experiments",
+        "text_level": 1,
+        "bbox": [
+            112,
+            549,
+            260,
+            565
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1 Settings",
+        "text_level": 1,
+        "bbox": [
+            112,
+            577,
+            223,
+            593
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Datasets We conduct experiments on five low-resource translation tasks, four of which are from the Global Voices datasets (Tiedemann, 2012; Khayrallah et al., 2020): Polish (Pl), Hungarian (Hu), Indonesian (Id), Catalan (Ca) to English (En), where we use the officially provided training sets, validation sets and test sets in our experiments. The other one is the WMT 2017 Turkish (Tr) to En benchmark. We use newstest2016 as the validation set and newstest2017 as the test set. For the parent models training, we use the German-English dataset following the empirical advice of (Aji et al., 2020; Li et al., 2022). We take the WMT 2017 news translation task as our parent dataset containing around 5.8M paired sentences. The detailed statistics of these parallel corpora are presented in Table 1. For fair comparisons, we adopt the same data preprocess techniques as previous research of TL (Li et al., 2022), which only apply normalization and tokenization to",
+        "bbox": [
+            112,
+            599,
+            489,
+            921
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/f335ad1e2a242402b7ad05670342c654f392620e7ce651d6e911ee3ea316402c.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Datasets</td><td># Train</td><td># Valid</td><td># Test</td></tr><tr><td>Global Voices PI - En</td><td>39.9K</td><td>2,000</td><td>2,000</td></tr><tr><td>Global Voices Ca -En</td><td>15.2K</td><td>2,000</td><td>2,000</td></tr><tr><td>Global Voices Id - En</td><td>8.4K</td><td>2,000</td><td>2,000</td></tr><tr><td>Global Voices Hu - En</td><td>7.7K</td><td>2,000</td><td>2,000</td></tr><tr><td>WMT 2017 Tr - En</td><td>196.6K</td><td>3,000</td><td>3,007</td></tr><tr><td>WMT 2017 De - En</td><td>5.8M</td><td>3,000</td><td>3,003</td></tr></table>",
+        "bbox": [
+            524,
+            82,
+            870,
+            198
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 1: The statistics of parallel corpora.",
+        "bbox": [
+            552,
+            206,
+            836,
+            221
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "parallel sentences by using Moses toolkit<sup>1</sup>. Further, we apply Byte Pair Encoding (BPE) (Sennrich et al., 2016) to address the out-of-vocabulary problem and segment words with 16,000 merge operations for Turkish and 8,000 for the rest.",
+        "bbox": [
+            507,
+            252,
+            882,
+            332
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Model Configuration In our experiments, we implement translation models with fairseq toolkit. We choose the Transformer (Vaswani et al., 2017) as the backbone to implement our framework. We use Transformer_base that consists of 6 encoder and decoder layers with 8 attention heads. The number of dimensions of all sub-layers in the model is set to 512, and the inner layers of feed-forward layers have 2048 dimensions. Our models are trained on 2 Nvidia A100 GPUs. We train our models using Adam (Kingma and Ba, 2015) with  $(\\beta_{1},\\beta_{2}) = (0.9,0.98)$  and use cross-entropy as criterion with label smoothing  $= 0.1$ . In addition, we train the forward and backward parent model (i.e.,  $\\mathrm{De}\\rightarrow \\mathrm{En}$  and  $\\mathrm{En}\\rightarrow \\mathrm{De}$ ) with the initial learning rate  $1e^{-7}$  and gradually increase till  $1e^{-3}$  within 10,000 warm-up updates. For the models with transfer learning, we set the initial learning rate to  $1e^{-7}$ , and the peak learning rate is  $2e^{-4}$  within 1,000 warm-up steps. Dropout is applied to the output of each sub-layer with a rate of 0.3 to avoid over-fitting. Besides, attention and activation dropouts are also used with a rate of 0.1 and 0.1. We train all models with a maximum of 200 epochs and select the checkpoints with the best BLEU score on the validation set as our final model, where beam search is applied with beam size 5, and the length penalty is 1.",
+        "bbox": [
+            507,
+            351,
+            884,
+            803
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Baselines We use the following baselines to validate our method:",
+        "bbox": [
+            507,
+            821,
+            882,
+            852
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "https://github.com/moses-smt/ mosesdecoder",
+        "bbox": [
+            507,
+            869,
+            867,
+            895
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{2}$ https://github.com/facebookresearch/fairseq",
+        "bbox": [
+            507,
+            895,
+            867,
+            921
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "3218",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/e317745a44271e1b70eb8bc923795d25255c4cf2f6b8b5e569d1e236da6065c3.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Model</td><td colspan=\"2\">Tr→En</td><td colspan=\"2\">Hu→En</td><td colspan=\"2\">Id→En</td><td colspan=\"2\">Ca→En</td><td colspan=\"2\">Pl→En</td></tr><tr><td>BLEU</td><td>BS</td><td>BLEU</td><td>BS</td><td>BLEU</td><td>BS</td><td>BLEU</td><td>BS</td><td>BLEU</td><td>BS</td></tr><tr><td>Vanilla</td><td>17.8</td><td>51.8</td><td>0.9</td><td>0.9</td><td>1.1</td><td>13.2</td><td>1.1</td><td>15.5</td><td>1.5</td><td>18.9</td></tr><tr><td>TL</td><td>17.6</td><td>51.9</td><td>5.9</td><td>27.4</td><td>13.5</td><td>37.7</td><td>21.6</td><td>51.8</td><td>19.9</td><td>55.3</td></tr><tr><td>TM-TL</td><td>18.6</td><td>53.9</td><td>10.6</td><td>41.2</td><td>18.6</td><td>49.9</td><td>25.3</td><td>58.9</td><td>21.4</td><td>58.2</td></tr><tr><td>ConsistTL</td><td>19.3</td><td>55.9</td><td>11.9</td><td>43.9</td><td>19.7</td><td>52.2</td><td>26.6</td><td>60.0</td><td>22.4</td><td>59.9</td></tr><tr><td>TSFT (ours)</td><td>20.0</td><td>56.7</td><td>13.1</td><td>44.6</td><td>20.5</td><td>53.3</td><td>27.7</td><td>60.7</td><td>23.3</td><td>60.5</td></tr></table>",
+        "bbox": [
+            119,
+            82,
+            878,
+            247
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Vanilla NMT (Vaswani et al., 2017): A bilingual NMT model with Transformer architecture directly trained on low-resource child training data from scratch.",
+            "- TL (Zoph et al., 2016): The first transfer learning work for NMT, initializing the child model with a parent model except for the source word embeddings. Note that the original work employed a two-layer encoder-decoder LSTM model, whereas we replicate TL using Transformer.",
+            "- TM-TL (Aji et al., 2020): To transfer embeddings across languages with distinct linguistic characteristics, Token Matching (TM) is proposed to assign the child word embeddings with the same tokens in the parent embeddings. The remaining unmatched tokens are assigned random embeddings as TL.",
+            "- ConsistTL (Li et al., 2022): Based on TM-TL, ConsistTL is proposed to enhance the child model by incorporating the prediction of the parent model during the fine-tuning of the child model."
+        ],
+        "bbox": [
+            136,
+            338,
+            489,
+            744
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Metrics To validate the effectiveness of our proposed framework, we use the following two metrics:",
+        "bbox": [
+            112,
+            758,
+            489,
+            806
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "- BLEU (Papineni et al., 2002): Considering the discrepancy among different tokenization processes, we apply the SacreBLEU score (Post, 2018)<sup>3</sup> for all experiments.",
+        "bbox": [
+            136,
+            820,
+            487,
+            885
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/34e16179efada132c18e2bbcc118a4ffdace5d6a49d9ad538213e41ab9977a09.jpg",
+        "table_caption": [
+            "Table 2: The SacreBLEU and BERTScore scores of baselines and ours on various translations. \"BS\" represents BERTScore. Blod indicates the best result. BLEU score reflects that TSFT is significantly better than ConsistTL with t-test  $p < 0.05$ . The number of bootstrap resamples is set to 1,000 to measure the significant difference between results."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Hyper-parameter</td><td>Tr→En</td><td>Hu→En</td></tr><tr><td>(λ = 2.0, τ = 2.0)</td><td>19.9</td><td>13.0</td></tr><tr><td>(λ = 3.0, τ = 2.0)</td><td>19.8</td><td>12.8</td></tr><tr><td>(λ = 4.0, τ = 2.0)</td><td>20.0</td><td>13.1</td></tr><tr><td>(λ = 5.0, τ = 2.0)</td><td>19.9</td><td>12.9</td></tr><tr><td>(λ = 4.0, τ = 0.5)</td><td>19.7</td><td>12.9</td></tr><tr><td>(λ = 4.0, τ = 1.0)</td><td>19.7</td><td>13.1</td></tr><tr><td>(λ = 4.0, τ = 3.0)</td><td>19.4</td><td>13.0</td></tr></table>",
+        "bbox": [
+            554,
+            336,
+            838,
+            463
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 3: The SacreBLEU scores on the test set of the  $\\mathrm{Tr} \\rightarrow \\mathrm{En}$  and  $\\mathrm{Hu} \\rightarrow \\mathrm{En}$  translations with different  $\\lambda$  and  $\\tau$ .",
+        "bbox": [
+            507,
+            472,
+            882,
+            501
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "- BERTScore (Zhang et al., 2020): Leveraging a pre-trained BERT model to evaluate the semantic correctness between the predictions and references by cosine similarity.",
+        "bbox": [
+            531,
+            536,
+            882,
+            601
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2 Main Results",
+        "text_level": 1,
+        "bbox": [
+            507,
+            633,
+            660,
+            646
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The results on five low-resource translation benchmarks are presented in Table 2. In our experiments, we utilize German as the parent language, and the parent models are pre-trained on a German-to-English dataset. As we can see, our method significantly outperforms the vanilla NMT in terms of both SacreBLEU and BERTScore. Compared with TL and TM-TL, TSFT still achieves significant improvements on all translations. Moreover, our proposed TSFT also has demonstrated superior performance compared to the strongest baseline ConsistTL with up to +1.2 SacreBLEU points and +1.1 BERTScore points. Overall, these results prove that our proposed transfer learning framework TSFT can effectively improve the performance of the child model on low-resource translation tasks.",
+        "bbox": [
+            505,
+            664,
+            884,
+            920
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^3$ Signature: nrefs:1 + case:mixed + eff:no + tok:13a + smooth:exp + version:2.0.0",
+        "bbox": [
+            112,
+            894,
+            489,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "3219",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/8651d18bfa80a35e1e777c1dab62df2914a824210edfdb0a544b9784964a2b44.jpg",
+        "image_caption": [
+            "Figure 3: The SacreBLEU scores of TSFT with different hyper-parameter  $l$  on  $\\mathrm{Tr} \\rightarrow \\mathrm{En}$  and  $\\mathrm{Hu} \\rightarrow \\mathrm{En}$ .  $\\mathrm{De} \\Rightarrow \\mathrm{Tr} / \\mathrm{Hu}$  indicates De is the parent language and  $\\mathrm{Tr} / \\mathrm{Hu}$  is the child language."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            122,
+            87,
+            487,
+            195
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/e457db89155d57fceb7e4aa6ab6cc179e6705f2369096830d37f54998f7a8791.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Models</td><td>Tr→En</td><td>Hu→En</td></tr><tr><td>TSFT</td><td>20.0</td><td>13.1</td></tr><tr><td>w/o PDF</td><td>19.5</td><td>12.5</td></tr><tr><td>w/o Ldist</td><td>19.8</td><td>12.8</td></tr><tr><td>w/o Step 2</td><td>18.9</td><td>11.2</td></tr><tr><td>w/o Step 2 + PDF</td><td>18.6</td><td>10.6</td></tr></table>",
+        "bbox": [
+            154,
+            281,
+            448,
+            380
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 4: The SacreBLEU scores on the test set of the  $\\mathrm{Tr}\\rightarrow \\mathrm{En}$  and  $\\mathrm{Hu}\\rightarrow \\mathrm{En}$  translations with PDF,  $L_{dist}$ , and Step 2 ablation.",
+        "bbox": [
+            112,
+            388,
+            487,
+            432
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5 Analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            456,
+            223,
+            473
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.1 Effect of the Number of Freezing Layers",
+        "text_level": 1,
+        "bbox": [
+            112,
+            481,
+            478,
+            498
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In Section 3.2, we utilize the PDF strategy in Step 1. However, we do not clearly know the optimal number of freezing layers  $l$  that can benefit the child model most. Different numbers of freezing layers would significantly impact the child model performance. Hence, in this section, we conduct a comparative analysis of the impact of different  $l$  on the translation performance of the child model.",
+        "bbox": [
+            112,
+            502,
+            485,
+            631
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Concretely, we still use the  $\\mathrm{De}\\rightarrow \\mathrm{En}$  model as the parent model and select  $\\mathrm{Tr}\\to \\mathrm{En}$  and  $\\mathrm{Hu}\\to \\mathrm{En}$  translations as child tasks. We tune the hyperparameter  $l$  by performing a grid search on  $l\\in$ $\\{1,2,3,4,5,6\\}$ . Figure 3 illustrates the model performance with different values of  $l$ . We can find that the final child models achieve the best performance in  $\\mathrm{Tr}\\to \\mathrm{En}$  and  $\\mathrm{Hu}\\to \\mathrm{En}$  when  $l$  is 5 and 4, respectively. Consequently, we set  $l$  as 5 for  $\\mathrm{Tr}\\to \\mathrm{En}$  translation and 4 for the rest.",
+        "bbox": [
+            112,
+            631,
+            487,
+            791
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Despite a substantial size difference between the  $\\mathrm{Tr} \\rightarrow \\mathrm{En}$  and  $\\mathrm{Hu} \\rightarrow \\mathrm{En}$  datasets, there is not much difference in the choice of the number of layers to freeze. For this phenomenon, we speculate that the distinction between these two child datasets is negligible compared to the size distinctions with the parent dataset, as shown in Table 1. Therefore, when applying our framework to parent models",
+        "bbox": [
+            112,
+            791,
+            487,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/4d7af7a37c77509d54c570ab7bd478f725e3e82560ba64a4c6f4c513e8506378.jpg",
+        "image_caption": [
+            "Figure 4: Learning curves of different TL methods."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            522,
+            90,
+            870,
+            269
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "with relatively limited resources, the choice of the number of frozen decoder layers needs to be carefully considered to achieve optimal results.",
+        "bbox": [
+            507,
+            328,
+            882,
+            376
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.2 Effect of Hyper-parameters  $\\lambda$  and  $\\tau$",
+        "text_level": 1,
+        "bbox": [
+            507,
+            386,
+            835,
+            401
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Hyper-parameter  $\\lambda$  is crucial to controlling the influence of the two losses within the  $L_{child}$ . In this part, we set  $\\lambda$  to  $\\{2.0, 3.0, 4.0, 5.0\\}$  to investigate the impact of different values of  $\\lambda$  on the performance of the child model. The corresponding SacreBLEU scores are presented in Table 3. For both  $\\mathrm{Tr} \\rightarrow \\mathrm{En}$  and  $\\mathrm{Hu} \\rightarrow \\mathrm{En}$  translations, the best performances are obtained when  $\\lambda$  is set to 4.0. Hence, we set  $\\lambda$  as 4.0 for all experiments involving  $L_{dist}$ .",
+        "bbox": [
+            507,
+            407,
+            882,
+            552
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In addition, we also conduct experiments with varying values of  $\\tau$  during the training process of the child model, while keeping  $\\lambda$  fixed at 4.0. As illustrated in Table 3, we can find that the performance of the child model is sensitive to  $\\tau$  and the performance is best when  $\\tau$  is set to 2.0. We argue that this is because minimizing the KL divergence is difficult, but using a larger  $\\tau$  (e.g., 3.0) may diminish the information from the intermediate model, which is not helpful in improving the performance of the child model.",
+        "bbox": [
+            507,
+            552,
+            882,
+            728
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.3 Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            507,
+            740,
+            673,
+            755
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We conduct an ablation study of the PDF strategy,  $L_{dist}$ , and Step 2 to explore their effects on our framework. We present the performance of four variants of TSFT as follows: 1) w/o PDF. During the training process of Step 1, we do not freeze any layers of the intermediate model, fine-tuning all parameters in every epoch. 2) w/o  $L_{dist}$ . In Step 2, we eliminate the distillation loss between the encoders of the intermediate and child models, conducting fine-tuning of the child model using",
+        "bbox": [
+            507,
+            760,
+            882,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "3220",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/5953f927bc5a3643ef28211f92a6b3339d9850b87ce090a6e3c24cf2a9d77c83.jpg",
+        "image_caption": [
+            "(a) TM-TL"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            206,
+            93,
+            386,
+            219
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/408affc16d105e3d1d6d5ad052c8b3fe4390fb1498d545a79a25756a1d67f8dd.jpg",
+        "image_caption": [
+            "(b) ConsistTL",
+            "Figure 5: Sentence representations after using T-SNE dimensionality reduction. The blue points denote the output from the parent model, and the red points denote the output from the fine-tuned models obtained from different transfer learning methods."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            405,
+            93,
+            589,
+            219
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/9f6ab930631684a907d33bea6ddcc6c2bf444f77e8e41fc23e9c4e8c4a2bfebf.jpg",
+        "image_caption": [
+            "(c) TSFT (Step 1)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            613,
+            93,
+            791,
+            218
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "$L_{ce}$  exclusively. 3) w/o Step 2. We evaluate the translation performance of the intermediate model. 4) w/o Step  $2 + \\mathrm{PDF}$ . Based on 3), we do not freeze any layers of the intermediate model during Step 1. We conduct experiments on  $\\mathrm{Tr} \\rightarrow \\mathrm{En}$  and  $\\mathrm{Hu} \\rightarrow \\mathrm{En}$  translations, which correspondingly represent the largest and smallest datasets among those applied in our main experiments. The results are shown in Table 4. It is evident that excluding the PDF strategy,  $L_{dist}$ , or Step 2 resulting in a deterioration of the translation quality, underscoring the efficacy of these components within TSFT. The experimental results show that PDF has a greater impact than  $L_{dist}$ . Further, we observe that PDF can effectively improve the translation performance of the intermediate model and benefit the child model. This observation shows that retaining the performance of the parent model is crucial for improving the performance of the child model.",
+        "bbox": [
+            112,
+            326,
+            489,
+            632
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.4 Comparison of Learning Curves",
+        "text_level": 1,
+        "bbox": [
+            112,
+            643,
+            415,
+            659
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "A learning curve represents a model's learning performance throughout the duration of training and is a widely employed diagnostic tool in machine learning (Kambhatla et al., 2022; Bao et al., 2023). In this section, we present the validation learning curve to assess the generalization capabilities of TM-TL, ConsistTL, and TSFT by using the Sacre-BLEU score as the criterion. Figure 4 illustrates the learning curves of child models trained with three transfer learning methods. Compared with TM-TL and ConsistTL, TSFT exhibits superior initial performance and convergence speed. Note that the TSFT curve delineates the performance of the model fine-tuned after Step 1. This observation emphasizes the effectiveness of fine-tuning the intermediate model in enhancing the final model's",
+        "bbox": [
+            112,
+            664,
+            489,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "performance, which can be attributed to the augmentation of adaptability to child data consequent to the fine-tuning process in Step 1. Besides, as the training progresses into the stable phase, we can find that the performance of the child model under the TSFT framework is consistently higher than that of TM-TL and ConsistTL. It is noteworthy that, similar to TM-TL and ConsistTL, TSFT does not utilize additional data or resources. Thus, the performance improvement of the child model can be attributed to the effectiveness of the pre-finetune process.",
+        "bbox": [
+            507,
+            326,
+            884,
+            521
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.5 Sentence Representation Visualization",
+        "text_level": 1,
+        "bbox": [
+            507,
+            533,
+            855,
+            550
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In our framework, the intermediate model is used to adjust the parent parameters to perform well when using child source sentences as input (Section 3.2). Thus, in this section, we visualize the target-side sentence representations of the De-En parent model and Hu-En models obtained from different transfer learning methods. We utilize the T-SNE method (Hinton and Roweis, 2002) to project the representations into a 2-dimensional space, as shown in Figure 5. This figure shows that TM-TL struggles to align the child representations with the parent representations. ConsistL slightly reduces the discrepancy between the parent and child representations, whereas the intermediate model from TSFT makes the representations much more similar. This observation shows that our fine-tuned intermediate model can produce similar outputs to the parent model even with different source languages.",
+        "bbox": [
+            507,
+            556,
+            884,
+            847
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            860,
+            640,
+            876
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we propose TSFT: a novel two-step fine-tuning framework for low-resource NMT.",
+        "bbox": [
+            507,
+            890,
+            884,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "3221",
+        "bbox": [
+            480,
+            927,
+            517,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "TSFT incorporates an intermediate (child) model to pre-fine-tune the parent model to fit the child data. The intermediate model is initialized with the parent model and then fine-tuned on the child source data in the first step. We propose freezing partial decoder layers when fine-tuning the intermediate model to alleviate catastrophic forgetting. In the second step, TSFT initializes the child model with the intermediate model and fine-tunes the child model on the parallel data using the cross-entropy and proposed distillation losses. Experimental results on five low-resource translations demonstrate the effectiveness of our proposed TSFT.",
+        "bbox": [
+            112,
+            84,
+            492,
+            294
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            112,
+            309,
+            220,
+            324
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "When using our proposed framework, two fine-tuning steps are necessary to obtain the final child model. Therefore, compared to one-step transfer learning methods in NMT, TSFT may require more training time and computation resources to transfer parent knowledge to the child model. Nevertheless, it is important to note that TSFT does not introduce additional time or computing resource consumption during inference. Besides, TSFT is designed for transfer learning scenarios when the target languages of the parent and child models are identical. We will try transferring different target languages in the future.",
+        "bbox": [
+            112,
+            338,
+            489,
+            546
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Ethics Statement",
+        "text_level": 1,
+        "bbox": [
+            114,
+            563,
+            265,
+            577
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This study uses only publicly accessible datasets and models that permit academic research. The preprocessing tools and model training toolkit are open-sourced without copyright conflicts.",
+        "bbox": [
+            112,
+            592,
+            487,
+            657
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            114,
+            671,
+            285,
+            688
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We would like to thank the anonymous reviewers for their helpful comments. This work is supported by the 2020 Catalyst: Strategic New Zealand - Singapore Data Science Research Programme Fund by Ministry of Business, Innovation and Employment (MBIE), New Zealand.",
+        "bbox": [
+            112,
+            701,
+            489,
+            797
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            828,
+            213,
+            843
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Roee Aharoni, Melvin Johnson, and Orhan First. 2019. Massively multilingual neural machine translation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics.",
+        "bbox": [
+            115,
+            853,
+            489,
+            921
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Alham Fikri Aji, Nikolay Bogoychev, Kenneth Heafield, and Rico Sennrich. 2020. In neural machine translation, what does transfer learning transfer? In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics.",
+        "bbox": [
+            510,
+            85,
+            884,
+            152
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Guangsheng Bao, Zhiyang Teng, and Yue Zhang. 2023. Target-side augmentation for document-level machine translation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics.",
+        "bbox": [
+            509,
+            162,
+            884,
+            231
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Alexandre Bérard. 2021. Continual learning in multilingual NMT via language-specific embeddings. In Proceedings of the Sixth Conference on Machine Translation.",
+        "bbox": [
+            509,
+            241,
+            884,
+            294
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Deng Cai, Yan Wang, Huayang Li, Wai Lam, and Lemao Liu. 2021. Neural machine translation with monolingual translation memory. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing.",
+        "bbox": [
+            509,
+            306,
+            882,
+            387
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Chenhui Chu, Raj Dabre, and Sadao Kurohashi. 2017. An empirical comparison of domain adaptation methods for neural machine translation. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics.",
+        "bbox": [
+            509,
+            398,
+            884,
+            464
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Juuso Eronen, Michal Ptaszynski, Karol Nowakowski, Zheng Lin Chia, and Fumito Masui. 2023. Improving polish to english neural machine translation with transfer learning: Effects of data volume and language similarity. arXiv preprint arXiv:2306.00660.",
+        "bbox": [
+            509,
+            476,
+            882,
+            543
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Naman Goyal, Cynthia Gao, Vishrav Chaudhary, Peng-Jen Chen, Guillaume Wenzek, Da Ju, Sanjana Krishnan, Marc'Aurelio Ranzato, Francisco Guzmán, and Angela Fan. 2022. The Flores-101 evaluation benchmark for low-resource and multilingual machine translation. Transactions of the Association for Computational Linguistics.",
+        "bbox": [
+            509,
+            554,
+            882,
+            646
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Shuhao Gu and Yang Feng. 2020. Investigating catastrophic forgetting during continual training for neural machine translation. In Proceedings of the 28th International Conference on Computational Linguistics.",
+        "bbox": [
+            509,
+            658,
+            884,
+            713
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Shuhao Gu, Yang Feng, and Wanying Xie. 2021. Pruning-then-expanding model for domain adaptation of neural machine translation. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics.",
+        "bbox": [
+            509,
+            722,
+            884,
+            790
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Dengji Guo, Zhengrui Ma, Min Zhang, and Yang Feng. 2022. Prediction difference regularization against perturbation for neural machine translation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics.",
+        "bbox": [
+            509,
+            802,
+            882,
+            869
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Geoffrey E Hinton and Sam Roweis. 2002. Stochastic neighbor embedding. Advances in neural information processing systems.",
+        "bbox": [
+            509,
+            879,
+            882,
+            921
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "3222",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Aizhan Imankulova, Raj Dabre, Atsushi Fujita, and Kenji Imamura. 2019. Exploiting out-of-domain parallel data through multilingual transfer learning for low-resource neural machine translation. In Proceedings of Machine Translation Summit XVII.",
+            "Baijun Ji, Zhirui Zhang, Xiangyu Duan, Min Zhang, Boxing Chen, and Weihua Luo. 2020. Cross-lingual pre-training based transfer for zero-shot neural machine translation. In Proceedings of the AAAI conference on artificial intelligence.",
+            "Nishant Kambhatla, Logan Born, and Anoop Sarkar. 2022. Ciphertext based data augmentation for neural machine translation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics.",
+            "Huda Khayrallah, Brian Thompson, Matt Post, and Philipp Koehn. 2020. Simulated multiple reference training improves low-resource machine translation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing.",
+            "Yunsu Kim, Yingbo Gao, and Hermann Ney. 2019a. Effective cross-lingual transfer of neural machine translation models without shared vocabularies. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics.",
+            "Yunsu Kim, Petre Petrov, Pavel Petrushkov, Shahram Khadivi, and Hermann Ney. 2019b. Pivot-based transfer learning for neural machine translation between non-english languages. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing.",
+            "Diederik P Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In Proceedings of the 3rd International Conference for Learning Representations.",
+            "James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. 2017. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences.",
+            "Tom Kocmi and Ondrej Bojar. 2018. Trivial transfer learning for low-resource neural machine translation. Proceedings of the Third Conference on Machine Translation: Research Papers.",
+            "Tom Kocmi and Ondrej Bojar. 2020. Efficiently reusing old models across languages via transfer learning. In Proceedings of the 22nd Annual Conference of the European Association for Machine Translation.",
+            "Surafel M Lakew, Alija Erofeeva, Matteo Negri, Marcello Federico, and Marco Turchi. 2018. Transfer learning in multilingual neural machine translation with dynamic vocabulary. In Proceedings of the 15th International Conference on Spoken Language Translation."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Zhaocong Li, Xuebo Liu, Derek F. Wong, Lidia S. Chao, and Min Zhang. 2022. ConsistTL: Modeling consistency in transfer learning for low-resource neural machine translation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing.",
+            "J. Lin. 1991. Divergence measures based on the shannon entropy. IEEE Transactions on Information Theory.",
+            "Yu-Hsiang Lin, Chian-Yu Chen, Jean Lee, Zirui Li, Yuyan Zhang, Mengzhou Xia, Shruti Rijhwani, Junxian He, Zhisong Zhang, Xuezhe Ma, et al. 2019. Choosing transfer languages for cross-lingual learning. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics.",
+            "Shudong Liu, Xuebo Liu, Derek F Wong, Zhaocong Li, Wenxiang Jiao, Lidia S Chao, and Min Zhang. 2023. kNN-TL: k-nearest-neighbor transfer learning for low-resource neural machine translation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics.",
+            "Gongxu Luo, Yating Yang, Yang Yuan, Zhanheng Chen, and Aizimaiti Ainiwaer. 2019. Hierarchical transfer learning architecture for low-resource neural machine translation. IEEE Access.",
+            "Mieradilijiang Maimaiti, Yang Liu, Huanbo Luan, and Maosong Sun. 2019. Multi-round transfer learning for low-resource nmt using multiple high-resource languages. ACM Transactions on Asian and Low-Resource Language Information Processing.",
+            "Paul Michel and Graham Neubig. 2018. Mtnt: A testbed for machine translation of noisy text. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing.",
+            "Toan Q Nguyen and David Chiang. 2017. Transfer learning across low-resource, related languages for neural machine translation. In Proceedings of the Eighth International Joint Conference on Natural Language Processing.",
+            "Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. BLEU: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics.",
+            "Peyman Passban, Qun Lui, and Andy Way. 2017. Translating low-resource languages by vocabulary adaptation from close counterparts. ACM Transactions on Asian and Low-Resource Language Information Processing.",
+            "Frithjof Petrick, Jan Rosendahl, Christian Herold, and Hermann Ney. 2022. Locality-sensitive hashing for long context neural machine translation. In Proceedings of the 19th International Conference on Spoken Language Translation."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            920
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "3223",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Matt Post. 2018. A call for clarity in reporting BLEU scores. In Proceedings of the Third Conference on Machine Translation.",
+            "Keisuke Sakaguchi, Kevin Duh, Matt Post, and Benjamin Van Durme. 2017. Robsut wrod reocginiton via semi-character recurrent neural network. In Thirty-first AAAI conference on artificial intelligence.",
+            "Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Neural machine translation of rare words with subword units. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics.",
+            "Hendra Setiawan, Zhongqiang Huang, and Rabih Zbib. 2018. BBN's low-resource machine translation for the LoReHLT 2016 evaluation. Machine Translation.",
+            "Brian Thompson, Jeremy G Winnup, Huda Khayrallah, Kevin Duh, and Philipp Koehn. 2019. Overcoming catastrophic forgetting during domain adaptation of neural machine translation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies.",
+            "Jörg Tiedemann. 2012. Parallel data, tools and interfaces in OPUS. In Proceedings of the Eighth International Conference on Language Resources and Evaluation.",
+            "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems.",
+            "Yuqiao Wen, Zichao Li, Wenyu Du, and Lili Mou. 2023. f-Divergence minimization for sequence-level knowledge distillation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics.",
+            "Mengzhou Xia, Xiang Kong, Antonios Anastasopoulos, and Graham Neubig. 2019. Generalized data augmentation for low-resource translation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics.",
+            "Pengcheng Yin, Hao Fang, Graham Neubig, Adam Pauls, Emmanouil Antonios Platanios, Yu Su, Sam Thomson, and Jacob Andreas. 2021. Compositional generalization for neural semantic parsing via span-level supervised attention. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics.",
+            "Tianyi Zhang, Varsha Kishore, Felix Wu0, Kilian Q. Weinberger, and Yoav Artzi. 2020. BERTScore: Evaluating text generation with BERT. In International Conference on Learning Representations.",
+            "Jiawei Zhou and Phillip Keung. 2020. Improving nonautoregressive neural machine translation with monolingual data. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics."
+        ],
+        "bbox": [
+            115,
+            85,
+            485,
+            919
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "ref_text",
+        "text": "Barret Zoph, Deniz Yuret, Jonathan May, and Kevin Knight. 2016. Transfer learning for low-resource neural machine translation. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing.",
+        "bbox": [
+            510,
+            85,
+            880,
+            151
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "3224",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 10
+    }
+]

2024/A Novel Two-step Fine-tuning Framework for Transfer Learning in Low-Resource Neural Machine Translation/07401a9f-9a22-4566-a6eb-41c799e41ff7_model.json

@@ -0,0 +1,2037 @@
+[
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.15,
+                0.09,
+                0.848,
+                0.131
+            ],
+            "angle": 0,
+            "content": "A Novel Two-step Fine-tuning Framework for Transfer Learning in Low-Resource Neural Machine Translation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.35,
+                0.156,
+                0.654,
+                0.173
+            ],
+            "angle": 0,
+            "content": "Yuan Gao, Feng Hou*, Ruili Wang"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.15,
+                0.175,
+                0.852,
+                0.193
+            ],
+            "angle": 0,
+            "content": "School of Mathematical and Computational Science, Massey University, New Zealand"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.266,
+                0.196,
+                0.738,
+                0.213
+            ],
+            "angle": 0,
+            "content": "{y.gao, f.hou, ruili.wang}@massey.ac.nz"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.261,
+                0.262,
+                0.341,
+                0.276
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.142,
+                0.284,
+                0.461,
+                0.596
+            ],
+            "angle": 0,
+            "content": "Existing transfer learning methods for neural machine translation typically use a well-trained translation model (i.e., a parent model) of a high-resource language pair to directly initialize a translation model (i.e., a child model) of a low-resource language pair, and the child model is then fine-tuned with corresponding datasets. In this paper, we propose a novel two-step fine-tuning (TSFT) framework for transfer learning in low-resource neural machine translation. In the first step, we adjust the parameters of the parent model to fit the child language by using the child source data. In the second step, we transfer the adjusted parameters to the child model and fine-tune it with a proposed distillation loss for efficient optimization. Our experimental results on five low-resource translations demonstrate that our framework yields significant improvements over various strong transfer learning baselines. Further analysis demonstrated the effectiveness of different components in our framework."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.604,
+                0.262,
+                0.619
+            ],
+            "angle": 0,
+            "content": "1 Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.629,
+                0.49,
+                0.869
+            ],
+            "angle": 0,
+            "content": "Neural machine translation (NMT) has achieved superior performance in terms of both fluency and adequacy for high-resource languages (Vaswani et al., 2017; Zhou and Keung, 2020; Cai et al., 2021; Guo et al., 2022). With the introduction of the attention mechanism (Yin et al., 2021; Petrick et al., 2022), NMT has been proven to be efficient and powerful in modeling long-distance dependencies. However, the performance of NMT systems deteriorates dramatically when insufficient parallel data are available for training (Sakaguchi et al., 2017; Michel and Neubig, 2018; Aharoni et al., 2019; Goyal et al., 2022). The scarcity of parallel corpora intensely limits the performance of an NMT system on low-resource languages."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.87,
+                0.49,
+                0.903
+            ],
+            "angle": 0,
+            "content": "Transfer learning is a learning paradigm for addressing the data scarcity problem (Zoph et al.,"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.541,
+                0.264,
+                0.865,
+                0.457
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.508,
+                0.471,
+                0.886,
+                0.528
+            ],
+            "angle": 0,
+            "content": "Figure 1: Comparison between vanilla transfer learning framework (a) and TSFT (b). Our proposed TSFT incorporates an intermediate model to pre-fine-tune the parent parameters to fit the child data."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.568,
+                0.886,
+                0.922
+            ],
+            "angle": 0,
+            "content": "2016; Nguyen and Chiang, 2017; Li et al., 2022). For NMT, transfer learning aims to transfer the knowledge from a well-trained high-resource translation model (i.e., a parent model, e.g., English \\(\\rightarrow\\) German) to a low-resource translation model (i.e., a child model, e.g., English \\(\\rightarrow\\) the Māori language). Prior transfer learning methods in NMT (Zoph et al., 2016; Chu et al., 2017) primarily achieve knowledge transfer by initializing the parameters of the child model with the parent model and fine-tuning the child model on the corresponding data. Such direct transfer of knowledge raises a vocabulary mismatch problem (Lakew et al., 2018; Lin et al., 2019; Kocmi and Bojar, 2020), and results in unsatisfied results for low-resource translations. Some methods have been proposed to alleviate the vocabulary mismatch problem, such as constructing joint dictionaries or employing a crosslingual token mapping technique (Passban et al., 2017; Kocmi and Bojar, 2018; Kim et al., 2019a). Additionally, Aji et al. (2020) proposed a token matching method that simply duplicates the embed"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.142,
+                0.907,
+                0.29,
+                0.921
+            ],
+            "angle": 0,
+            "content": "*Corresponding author."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3214"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.22,
+                0.946,
+                0.779,
+                0.973
+            ],
+            "angle": 0,
+            "content": "Findings of the Association for Computational Linguistics: NAACL 2024, pages 3214-3224 June 16-21, 2024 ©2024 Association for Computational Linguistics"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.486,
+                0.116
+            ],
+            "angle": 0,
+            "content": "dings of overlapping tokens from the parent model to the child model."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.118,
+                0.489,
+                0.31
+            ],
+            "angle": 0,
+            "content": "Recently, based on the work of Aji et al. (2020), Li et al. (2022) proposed ConsistTL that uses the predictions of the parent model to continuously provide soft targets during the fine-tuning of the child model. However, given the differences between the source inputs of the parent and the child translation tasks, the parent model is not an optimal starting point for the single-step fine-tuning of the child model using limited parallel child data. Therefore, it is necessary to pre-fine-tune the parent model to fit the child language before initializing the child model with it."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.312,
+                0.49,
+                0.81
+            ],
+            "angle": 0,
+            "content": "Building upon this insight, we propose a simple yet effective transfer learning framework, named Two-Step Fine-Tuning (TSFT), for low-resource NMT. As shown in Figure 1, we introduce an intermediate (child) model initialized with the parent model to adjust the parent parameters to fit the child language. TSFT involves two fine-tuning steps. In the first step, we feed child source sentences (i.e., monolingual data) and meaning-matched sentences in the parent source language into the intermediate and the parent models, respectively. Then, the intermediate model is fine-tuned with the objective of aligning probability distributions from the parent and intermediate models, aiming to adjust the parameters transferred from the parent model to perform well with child source sentences. Additionally, we propose a regularization-based strategy that can improve the translation performance of the intermediate model and benefit the child model. Note that we apply the token matching method to alleviate the vocabulary mismatch problem in the first step. In the second step, we transfer the adjusted parameters from the intermediate model to the child model and fine-tune the entire child model on the pertinent parallel data, employing both a cross-entropy loss and a proposed distillation loss. Extensive experiments on five low-resource translations show that TSFT surpasses the strongest baseline method with up to 1.2 SacreBLEU points. The ablation study demonstrates the effectiveness of different components within TSFT."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.132,
+                0.812,
+                0.489,
+                0.827
+            ],
+            "angle": 0,
+            "content": "Our contributions can be summarized as follows:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.136,
+                0.842,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": "- We propose a novel two-step fine-tuning framework for low-resource NMT, which introduces an intermediate (child) model to fit parent parameters for the data of child languages before initializing the child model with"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.546,
+                0.085,
+                0.68,
+                0.101
+            ],
+            "angle": 0,
+            "content": "the parent model."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.532,
+                0.114,
+                0.883,
+                0.162
+            ],
+            "angle": 0,
+            "content": "- We propose a regularization-based strategy for fine-tuning the intermediate model and a distillation loss for fine-tuning the child model."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.532,
+                0.174,
+                0.885,
+                0.238
+            ],
+            "angle": 0,
+            "content": "- We validate our method by extensive experiments on various low-resource translations and achieve improved performance compared to various transfer learning methods."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.532,
+                0.114,
+                0.885,
+                0.238
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.251,
+                0.662,
+                0.266
+            ],
+            "angle": 0,
+            "content": "2 Related work"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.277,
+                0.885,
+                0.76
+            ],
+            "angle": 0,
+            "content": "Existing studies have demonstrated the success of transfer learning for low-resource NMT (Lin et al., 2019; Imankulova et al., 2019; Ji et al., 2020; Eronen et al., 2023). Zoph et al. (2016) first introduced transfer learning into the field of NMT and proposed a parent-child framework, where parameters from a pre-trained parent model are directly transferred to a new child model with a shared target language. Subsequent research largely builds upon the parent-child framework and tends to leverage highly related parent language to perform transfer learning (Passban et al., 2017; Setiawan et al., 2018). However, the languages closely related to low-resource languages are also low-resourced (Nguyen and Chiang, 2017; Xia et al., 2019) and offer only modest performance improvements. Thus, researchers focused on identifying the critical factors for the effectiveness of the parent language. Experimental results from (Lin et al., 2019; Aji et al., 2020) emphasized that linguistic or geographical distance does not appear as important as the size of the parent data (Lin et al., 2019; Aji et al., 2020). This insight expands the range of parent languages available for transfer learning, and alleviates the limitations of highly related parent languages. Consequently, later researchers shifted their attention to parent languages with low relatedness but high-resourced. However, this exacerbates the vocabulary mismatch problem, posing a new challenge to transfer learning."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.761,
+                0.887,
+                0.921
+            ],
+            "angle": 0,
+            "content": "One solution to the vocabulary mismatch problem is to build a joint dictionary before training a parent model (Kocmi and Bojar, 2018; Kim et al., 2019b). However, this restricts the applicability of a pre-trained parent model to a specific child model only. To overcome this limitation, Kim et al. (2019a) proposed pre-training a language-agnostic cross-lingual word embedding independently from the parent model. Concurrently, token matching methods also show their effectiveness in transfer"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3215"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.49,
+                0.213
+            ],
+            "angle": 0,
+            "content": "learning without requiring additional training efforts (Aji et al., 2020; Kocmi and Bojar, 2020). Some other methods introduce highly related intermediate languages to gradually narrow the vocabulary disparity (Luo et al., 2019; Maimaiti et al., 2019). These methods take advantage of both large-scale data sources and syntactic similarity in the intermediate language."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.215,
+                0.49,
+                0.438
+            ],
+            "angle": 0,
+            "content": "Recently, Li et al. (2022) incorporated the idea of consistency learning into transfer learning based on the work of Aji et al. (2020) and proposed a novel transfer learning method called ConsistTL. This method enables the child model to utilize the parent model during fine-tuning. Subsequently, Liu et al. (2023) proposed kNN-TL, which extends ConsistTL by integrating a k-nearest neighbor (kNN) module, allowing the child model to utilize the parent model during inference. While our method also builds on ConsistTL, we focus on enhancing the child model's performance during fine-tuning. Thus, our work is orthogonal to kNN-TL."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.451,
+                0.218,
+                0.466
+            ],
+            "angle": 0,
+            "content": "3 Method"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.476,
+                0.489,
+                0.54
+            ],
+            "angle": 0,
+            "content": "In this section, we begin by providing an overview of the basic concepts behind transfer learning and then present our transfer learning framework, TSFT, in detail."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.551,
+                0.377,
+                0.567
+            ],
+            "angle": 0,
+            "content": "3.1 Transfer Learning Primary"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.572,
+                0.49,
+                0.812
+            ],
+            "angle": 0,
+            "content": "Given a source sentence \\( x = \\{x_{1},\\ldots ,x_{I}\\} \\), the objective of an NMT model is to translate it to a new sentence \\( y = \\{y_{1},\\dots ,y_{J}\\} \\) in a target language, where the source sentence and target sentence have lengths \\( I \\) and \\( J \\), respectively. A typical NMT model is composed of an encoder and a decoder. The encoder is designed to extract high-level semantic information from the source sentences and represent them as hidden states \\( H_{e} \\). The decoder generates the output probability \\( P(y_{i}|H_{e},y_{< i}) \\) of the next target token \\( y_{i} \\). An NMT model is trained on a parallel corpus by minimizing the cross-entropy (CE) loss between the predicted sentence and the ground-truth translation as follows:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.178,
+                0.825,
+                0.488,
+                0.868
+            ],
+            "angle": 0,
+            "content": "\\[\nL _ {c e} = - \\sum_ {i = 1} ^ {J} \\log P \\left(y _ {i} \\mid y _ {<   i}, x, \\theta\\right), \\tag {1}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.874,
+                0.489,
+                0.889
+            ],
+            "angle": 0,
+            "content": "where \\(\\theta\\) is the parameters of the entire NMT model."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.89,
+                0.489,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Transfer learning has been widely used when only limited training datasets are available for the"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.885,
+                0.31
+            ],
+            "angle": 0,
+            "content": "problem at hand. It transfers the knowledge acquired from large-scale data to enhance the model performance under low-resource conditions. Transfer learning typically follows a parent-child framework (Zoph et al., 2016), where it involves reusing the parameters \\(\\theta_{p}\\) from a pre-trained parent model to initialize part or all parameters of a child model. In the field of NMT, the parent model \\(\\mathcal{M}_p\\) is initially trained on a high-resourced parallel dataset \\(D_{p} = \\{X_{p},Y_{p}\\}\\), while there is only a limited-sized dataset \\(D_{c} = \\{X_{c},Y_{c}\\}\\) available to the child model \\(\\mathcal{M}_c\\). After the initialization step, the child model can be fine-tuned on \\(D_{c}\\), which is also optimized through the minimization of the CE loss."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.321,
+                0.725,
+                0.337
+            ],
+            "angle": 0,
+            "content": "3.2 Two-step Fine-tuning"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.342,
+                0.884,
+                0.47
+            ],
+            "angle": 0,
+            "content": "For NMT, an ideal transfer learning framework should enable the parent model to exert its complete capabilities on the child task. However, owing to the disparities between the parent and child languages, the current one-step fine-tuning transfer learning framework struggles to adjust the parameters of the parent model to fit the child source language under the constraints of limited child data."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.471,
+                0.884,
+                0.697
+            ],
+            "angle": 0,
+            "content": "The idea of TSFT is simple: before initializing the child model with the parent model, we first adjust the parameters of the parent model to enhance its congruity with the child source language. In this work, we propose to introduce an intermediate model, denoted as \\(\\mathcal{M}_a\\), to make the parameters of the parent model fit for the child data. Specifically, we initialize the intermediate model with the parent model and pre-fine-tune it by using the source side sentences of the child data, then fine-tune the child model with both the source and target child training data. Therefore, we design TSFT as a two-step framework, as shown in Figure 2."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.713,
+                0.885,
+                0.905
+            ],
+            "angle": 0,
+            "content": "Step 1: Intermediate Fine-tuning After initializing the intermediate model with a well-trained parent model, we aim to equip the intermediate model with the ability to utilize child source sentences as input for target language generation. Since the intermediate model and the parent model share the same target language, it is crucial to retain the generation ability of the parent model. Therefore, we input the source-side sentences of the child data to the intermediate model and the parent model and utilize the predicted distribution of the parent model as the soft label for fine-tuning."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.527,
+                0.906,
+                0.882,
+                0.921
+            ],
+            "angle": 0,
+            "content": "However, it is infeasible to directly input child"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3216"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.118,
+                0.083,
+                0.885,
+                0.273
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.281,
+                0.885,
+                0.34
+            ],
+            "angle": 0,
+            "content": "Figure 2: Our proposed transfer learning framework TSFT for low-resource NMT. In Step 1, the loss function \\( L_{inter} \\) is used to optimize the intermediate model. In Step 2, the child model is optimized by \\( L_{child} \\). The blue icy blocks are initialized with the parent model and frozen. The input German sentences are produced through back-translation."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.364,
+                0.49,
+                0.641
+            ],
+            "angle": 0,
+            "content": "source sentences into the parent model, given that the parent and child models have different source languages. Thus, we need a meaning-matched sentence for each child source sentence in the parent source language. In the context of low-resource translations, parallel data for non-English-centric is often limited in size or entirely absent, making it difficult to meet the requirements for intermediate fine-tuning. Therefore, we adopt the method of Li et al. (2022) to generate pseudo parent data \\( D_{p*} = \\{X_{p*}, Y_c\\} \\) by using a reversed parent model, where each \\( x_{p*} \\in X_{p*} \\) is aligned with \\( y_c \\in Y_c \\). Although such a method requires training a reverse parent model, it effectively generates meaning-matched input sentences for the parent model. In addition, we use the following loss function to optimize the intermediate model:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.138,
+                0.661,
+                0.489,
+                0.705
+            ],
+            "angle": 0,
+            "content": "\\[\nL _ {i n t e r} = \\sum_ {i = 1} ^ {J} F _ {d} [ P _ {i n t e r} (y _ {i}), P _ {p a r e n t} (y _ {i}) ], \\quad (2)\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.713,
+                0.49,
+                0.875
+            ],
+            "angle": 0,
+            "content": "where \\( F_{d} \\) is a distribution measurement method, in this work, we choose Jensen-Shannon (JS) divergence (Lin, 1991; Wen et al., 2023) as our \\( F_{d} \\). Our preliminary experiments find that JS divergence outperforms using Kullback-Leibler (KL) divergence when taking \\( P_{inter}(y_i) \\) as the first item and \\( P_{parent}(y_i) \\) as the second one. \\( P_{*}(y_i) \\) represents the prediction distributions of translation models at time step \\( i \\), which is conditioned on the input sentence and the previous tokens:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.21,
+                0.904,
+                0.488,
+                0.922
+            ],
+            "angle": 0,
+            "content": "\\[\nP _ {*} (y _ {i}) = P _ {*} (y _ {i} | x, y _ {<   i}). \\tag {3}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.364,
+                0.885,
+                0.462
+            ],
+            "angle": 0,
+            "content": "Before fine-tuning the intermediate model, we first apply the token matching method (Aji et al., 2020) that duplicates the embeddings of overlapping tokens from the parent and child vocabularies to alleviate the vocabulary mismatch problem."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.477,
+                0.887,
+                0.687
+            ],
+            "angle": 0,
+            "content": "Step 2: Child Fine-tuning In the second step, we employ the target-side sentences from the child training data as labels to fine-tune the child model with CE loss, following the general process of transfer learning. Since the encoder of the intermediate model has fine-tuned with the child source sentences, we argue that it encompasses valuable information that can facilitate the child model. Therefore, we extract the encoder outputs, \\( P_{*}^{e}(\\cdot) \\), from both the intermediate and child models and incorporate a distillation loss \\( L_{dist} \\) as an extra objective to optimize the child model by minimizing the KL divergence between two output representations:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.534,
+                0.71,
+                0.884,
+                0.753
+            ],
+            "angle": 0,
+            "content": "\\[\nL _ {d i s t} = - \\sum_ {i = 1} ^ {I} P _ {i n t e r} ^ {e} (x _ {i}) \\cdot l o g P _ {c h i l d} ^ {e} (x _ {i}), \\quad (4)\n\\]"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.586,
+                0.776,
+                0.882,
+                0.833
+            ],
+            "angle": 0,
+            "content": "\\[\n\\begin{array}{l} P _ {*} ^ {e} (x _ {i}) = P _ {*} ^ {e} (x _ {i} | x, \\tau) \\\\ = \\frac {\\exp \\left(z _ {i} / \\tau\\right)}{\\sum_ {j \\in V} \\exp \\left(z _ {j} / \\tau\\right)}, \\tag {5} \\\\ \\end{array}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.842,
+                0.885,
+                0.922
+            ],
+            "angle": 0,
+            "content": "where \\( I \\) denotes the sentence length of a child source sentence, \\( z \\) denotes the logits output of encoders before log softmax is computed, \\( V \\) represents the vocabulary, and \\( \\tau \\) is a temperate factor used to smooth the prediction distributions. As we"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3217"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.49,
+                0.15
+            ],
+            "angle": 0,
+            "content": "only reuse the output of encoders, the process of encoder distillation does not add any extra parameters to models. The overall loss is obtained by a weighted sum of \\( L_{ce} \\) and \\( L_{dist} \\):"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.213,
+                0.169,
+                0.489,
+                0.186
+            ],
+            "angle": 0,
+            "content": "\\[\nL _ {\\text {c h i l d}} = L _ {c e} + \\lambda L _ {\\text {d i s t}}, \\tag {6}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.197,
+                0.414,
+                0.213
+            ],
+            "angle": 0,
+            "content": "where \\(\\lambda\\) is a balancing hyper-parameter."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.23,
+                0.49,
+                0.535
+            ],
+            "angle": 0,
+            "content": "Partial Decoder Freeze Regularization-based methods are widely used to alleviate the catastrophic forgetting issue (Kirkpatrick et al., 2017; Gu and Feng, 2020; Gu et al., 2021). While updating all parameters typically yields good results on a new domain, the data distribution difference between the old and new domains can engender the issue of catastrophic forgetting, causing the fine-tuned model to abandon linguistic knowledge learned from previous dataset (Thompson et al., 2019; Bérard, 2021). In this work, we are interested in introducing the regularization-based technique during Step 1 to preserve the predictive capabilities of the parent model. We propose a Partial Decoder Freeze (PDF) strategy to freeze the parameters of the last \\(l\\) decoder layers of the intermediate model and only update the rest parameters. For the selection of parameters \\(l\\), we conducted empirical experiments in Section 5.1."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.55,
+                0.262,
+                0.567
+            ],
+            "angle": 0,
+            "content": "4 Experiments"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.578,
+                0.225,
+                0.594
+            ],
+            "angle": 0,
+            "content": "4.1 Settings"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.6,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Datasets We conduct experiments on five low-resource translation tasks, four of which are from the Global Voices datasets (Tiedemann, 2012; Khayrallah et al., 2020): Polish (Pl), Hungarian (Hu), Indonesian (Id), Catalan (Ca) to English (En), where we use the officially provided training sets, validation sets and test sets in our experiments. The other one is the WMT 2017 Turkish (Tr) to En benchmark. We use newstest2016 as the validation set and newstest2017 as the test set. For the parent models training, we use the German-English dataset following the empirical advice of (Aji et al., 2020; Li et al., 2022). We take the WMT 2017 news translation task as our parent dataset containing around 5.8M paired sentences. The detailed statistics of these parallel corpora are presented in Table 1. For fair comparisons, we adopt the same data preprocess techniques as previous research of TL (Li et al., 2022), which only apply normalization and tokenization to"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.525,
+                0.083,
+                0.871,
+                0.199
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Datasets</td><td># Train</td><td># Valid</td><td># Test</td></tr><tr><td>Global Voices PI - En</td><td>39.9K</td><td>2,000</td><td>2,000</td></tr><tr><td>Global Voices Ca -En</td><td>15.2K</td><td>2,000</td><td>2,000</td></tr><tr><td>Global Voices Id - En</td><td>8.4K</td><td>2,000</td><td>2,000</td></tr><tr><td>Global Voices Hu - En</td><td>7.7K</td><td>2,000</td><td>2,000</td></tr><tr><td>WMT 2017 Tr - En</td><td>196.6K</td><td>3,000</td><td>3,007</td></tr><tr><td>WMT 2017 De - En</td><td>5.8M</td><td>3,000</td><td>3,003</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.553,
+                0.208,
+                0.838,
+                0.222
+            ],
+            "angle": 0,
+            "content": "Table 1: The statistics of parallel corpora."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.253,
+                0.884,
+                0.334
+            ],
+            "angle": 0,
+            "content": "parallel sentences by using Moses toolkit<sup>1</sup>. Further, we apply Byte Pair Encoding (BPE) (Sennrich et al., 2016) to address the out-of-vocabulary problem and segment words with 16,000 merge operations for Turkish and 8,000 for the rest."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.353,
+                0.885,
+                0.804
+            ],
+            "angle": 0,
+            "content": "Model Configuration In our experiments, we implement translation models with fairseq toolkit. We choose the Transformer (Vaswani et al., 2017) as the backbone to implement our framework. We use Transformer_base that consists of 6 encoder and decoder layers with 8 attention heads. The number of dimensions of all sub-layers in the model is set to 512, and the inner layers of feed-forward layers have 2048 dimensions. Our models are trained on 2 Nvidia A100 GPUs. We train our models using Adam (Kingma and Ba, 2015) with \\((\\beta_{1},\\beta_{2}) = (0.9,0.98)\\) and use cross-entropy as criterion with label smoothing \\(= 0.1\\). In addition, we train the forward and backward parent model (i.e., \\(\\mathrm{De}\\rightarrow \\mathrm{En}\\) and \\(\\mathrm{En}\\rightarrow \\mathrm{De}\\)) with the initial learning rate \\(1e^{-7}\\) and gradually increase till \\(1e^{-3}\\) within 10,000 warm-up updates. For the models with transfer learning, we set the initial learning rate to \\(1e^{-7}\\), and the peak learning rate is \\(2e^{-4}\\) within 1,000 warm-up steps. Dropout is applied to the output of each sub-layer with a rate of 0.3 to avoid over-fitting. Besides, attention and activation dropouts are also used with a rate of 0.1 and 0.1. We train all models with a maximum of 200 epochs and select the checkpoints with the best BLEU score on the validation set as our final model, where beam search is applied with beam size 5, and the length penalty is 1."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.822,
+                0.884,
+                0.853
+            ],
+            "angle": 0,
+            "content": "Baselines We use the following baselines to validate our method:"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.509,
+                0.87,
+                0.868,
+                0.896
+            ],
+            "angle": 0,
+            "content": "https://github.com/moses-smt/ mosesdecoder"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.509,
+                0.896,
+                0.868,
+                0.922
+            ],
+            "angle": 0,
+            "content": "\\(^{2}\\)https://github.com/facebookresearch/fairseq"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.509,
+                0.87,
+                0.868,
+                0.922
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3218"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.12,
+                0.083,
+                0.88,
+                0.248
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Model</td><td colspan=\"2\">Tr→En</td><td colspan=\"2\">Hu→En</td><td colspan=\"2\">Id→En</td><td colspan=\"2\">Ca→En</td><td colspan=\"2\">Pl→En</td></tr><tr><td>BLEU</td><td>BS</td><td>BLEU</td><td>BS</td><td>BLEU</td><td>BS</td><td>BLEU</td><td>BS</td><td>BLEU</td><td>BS</td></tr><tr><td>Vanilla</td><td>17.8</td><td>51.8</td><td>0.9</td><td>0.9</td><td>1.1</td><td>13.2</td><td>1.1</td><td>15.5</td><td>1.5</td><td>18.9</td></tr><tr><td>TL</td><td>17.6</td><td>51.9</td><td>5.9</td><td>27.4</td><td>13.5</td><td>37.7</td><td>21.6</td><td>51.8</td><td>19.9</td><td>55.3</td></tr><tr><td>TM-TL</td><td>18.6</td><td>53.9</td><td>10.6</td><td>41.2</td><td>18.6</td><td>49.9</td><td>25.3</td><td>58.9</td><td>21.4</td><td>58.2</td></tr><tr><td>ConsistTL</td><td>19.3</td><td>55.9</td><td>11.9</td><td>43.9</td><td>19.7</td><td>52.2</td><td>26.6</td><td>60.0</td><td>22.4</td><td>59.9</td></tr><tr><td>TSFT (ours)</td><td>20.0</td><td>56.7</td><td>13.1</td><td>44.6</td><td>20.5</td><td>53.3</td><td>27.7</td><td>60.7</td><td>23.3</td><td>60.5</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.113,
+                0.256,
+                0.885,
+                0.314
+            ],
+            "angle": 0,
+            "content": "Table 2: The SacreBLEU and BERTScore scores of baselines and ours on various translations. \"BS\" represents BERTScore. Blod indicates the best result. BLEU score reflects that TSFT is significantly better than ConsistTL with t-test \\( p < 0.05 \\). The number of bootstrap resamples is set to 1,000 to measure the significant difference between results."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.339,
+                0.489,
+                0.404
+            ],
+            "angle": 0,
+            "content": "- Vanilla NMT (Vaswani et al., 2017): A bilingual NMT model with Transformer architecture directly trained on low-resource child training data from scratch."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.416,
+                0.49,
+                0.526
+            ],
+            "angle": 0,
+            "content": "- TL (Zoph et al., 2016): The first transfer learning work for NMT, initializing the child model with a parent model except for the source word embeddings. Note that the original work employed a two-layer encoder-decoder LSTM model, whereas we replicate TL using Transformer."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.54,
+                0.49,
+                0.653
+            ],
+            "angle": 0,
+            "content": "- TM-TL (Aji et al., 2020): To transfer embeddings across languages with distinct linguistic characteristics, Token Matching (TM) is proposed to assign the child word embeddings with the same tokens in the parent embeddings. The remaining unmatched tokens are assigned random embeddings as TL."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.665,
+                0.49,
+                0.745
+            ],
+            "angle": 0,
+            "content": "- ConsistTL (Li et al., 2022): Based on TM-TL, ConsistTL is proposed to enhance the child model by incorporating the prediction of the parent model during the fine-tuning of the child model."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.137,
+                0.339,
+                0.49,
+                0.745
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.759,
+                0.49,
+                0.807
+            ],
+            "angle": 0,
+            "content": "Metrics To validate the effectiveness of our proposed framework, we use the following two metrics:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.821,
+                0.489,
+                0.886
+            ],
+            "angle": 0,
+            "content": "- BLEU (Papineni et al., 2002): Considering the discrepancy among different tokenization processes, we apply the SacreBLEU score (Post, 2018)<sup>3</sup> for all experiments."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.556,
+                0.337,
+                0.839,
+                0.464
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Hyper-parameter</td><td>Tr→En</td><td>Hu→En</td></tr><tr><td>(λ = 2.0, τ = 2.0)</td><td>19.9</td><td>13.0</td></tr><tr><td>(λ = 3.0, τ = 2.0)</td><td>19.8</td><td>12.8</td></tr><tr><td>(λ = 4.0, τ = 2.0)</td><td>20.0</td><td>13.1</td></tr><tr><td>(λ = 5.0, τ = 2.0)</td><td>19.9</td><td>12.9</td></tr><tr><td>(λ = 4.0, τ = 0.5)</td><td>19.7</td><td>12.9</td></tr><tr><td>(λ = 4.0, τ = 1.0)</td><td>19.7</td><td>13.1</td></tr><tr><td>(λ = 4.0, τ = 3.0)</td><td>19.4</td><td>13.0</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.509,
+                0.473,
+                0.883,
+                0.502
+            ],
+            "angle": 0,
+            "content": "Table 3: The SacreBLEU scores on the test set of the \\(\\mathrm{Tr} \\rightarrow \\mathrm{En}\\) and \\(\\mathrm{Hu} \\rightarrow \\mathrm{En}\\) translations with different \\(\\lambda\\) and \\(\\tau\\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.532,
+                0.537,
+                0.884,
+                0.602
+            ],
+            "angle": 0,
+            "content": "- BERTScore (Zhang et al., 2020): Leveraging a pre-trained BERT model to evaluate the semantic correctness between the predictions and references by cosine similarity."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.634,
+                0.661,
+                0.648
+            ],
+            "angle": 0,
+            "content": "4.2 Main Results"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.665,
+                0.885,
+                0.921
+            ],
+            "angle": 0,
+            "content": "The results on five low-resource translation benchmarks are presented in Table 2. In our experiments, we utilize German as the parent language, and the parent models are pre-trained on a German-to-English dataset. As we can see, our method significantly outperforms the vanilla NMT in terms of both SacreBLEU and BERTScore. Compared with TL and TM-TL, TSFT still achieves significant improvements on all translations. Moreover, our proposed TSFT also has demonstrated superior performance compared to the strongest baseline ConsistTL with up to +1.2 SacreBLEU points and +1.1 BERTScore points. Overall, these results prove that our proposed transfer learning framework TSFT can effectively improve the performance of the child model on low-resource translation tasks."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.114,
+                0.895,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": "\\(^3\\)Signature: nrefs:1 + case:mixed + eff:no + tok:13a + smooth:exp + version:2.0.0"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3219"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.124,
+                0.089,
+                0.488,
+                0.196
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.112,
+                0.212,
+                0.49,
+                0.271
+            ],
+            "angle": 0,
+            "content": "Figure 3: The SacreBLEU scores of TSFT with different hyper-parameter \\( l \\) on \\( \\mathrm{Tr} \\rightarrow \\mathrm{En} \\) and \\( \\mathrm{Hu} \\rightarrow \\mathrm{En} \\). \\( \\mathrm{De} \\Rightarrow \\mathrm{Tr} / \\mathrm{Hu} \\) indicates De is the parent language and \\( \\mathrm{Tr} / \\mathrm{Hu} \\) is the child language."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.155,
+                0.282,
+                0.449,
+                0.381
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Models</td><td>Tr→En</td><td>Hu→En</td></tr><tr><td>TSFT</td><td>20.0</td><td>13.1</td></tr><tr><td>w/o PDF</td><td>19.5</td><td>12.5</td></tr><tr><td>w/o Ldist</td><td>19.8</td><td>12.8</td></tr><tr><td>w/o Step 2</td><td>18.9</td><td>11.2</td></tr><tr><td>w/o Step 2 + PDF</td><td>18.6</td><td>10.6</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.113,
+                0.389,
+                0.489,
+                0.434
+            ],
+            "angle": 0,
+            "content": "Table 4: The SacreBLEU scores on the test set of the \\(\\mathrm{Tr}\\rightarrow \\mathrm{En}\\) and \\(\\mathrm{Hu}\\rightarrow \\mathrm{En}\\) translations with PDF, \\(L_{dist}\\), and Step 2 ablation."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.457,
+                0.224,
+                0.474
+            ],
+            "angle": 0,
+            "content": "5 Analysis"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.482,
+                0.479,
+                0.499
+            ],
+            "angle": 0,
+            "content": "5.1 Effect of the Number of Freezing Layers"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.503,
+                0.487,
+                0.632
+            ],
+            "angle": 0,
+            "content": "In Section 3.2, we utilize the PDF strategy in Step 1. However, we do not clearly know the optimal number of freezing layers \\( l \\) that can benefit the child model most. Different numbers of freezing layers would significantly impact the child model performance. Hence, in this section, we conduct a comparative analysis of the impact of different \\( l \\) on the translation performance of the child model."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.632,
+                0.489,
+                0.792
+            ],
+            "angle": 0,
+            "content": "Concretely, we still use the \\(\\mathrm{De}\\rightarrow \\mathrm{En}\\) model as the parent model and select \\(\\mathrm{Tr}\\to \\mathrm{En}\\) and \\(\\mathrm{Hu}\\to \\mathrm{En}\\) translations as child tasks. We tune the hyperparameter \\(l\\) by performing a grid search on \\(l\\in\\) \\(\\{1,2,3,4,5,6\\}\\). Figure 3 illustrates the model performance with different values of \\(l\\). We can find that the final child models achieve the best performance in \\(\\mathrm{Tr}\\to \\mathrm{En}\\) and \\(\\mathrm{Hu}\\to \\mathrm{En}\\) when \\(l\\) is 5 and 4, respectively. Consequently, we set \\(l\\) as 5 for \\(\\mathrm{Tr}\\to \\mathrm{En}\\) translation and 4 for the rest."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.793,
+                0.489,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Despite a substantial size difference between the \\(\\mathrm{Tr} \\rightarrow \\mathrm{En}\\) and \\(\\mathrm{Hu} \\rightarrow \\mathrm{En}\\) datasets, there is not much difference in the choice of the number of layers to freeze. For this phenomenon, we speculate that the distinction between these two child datasets is negligible compared to the size distinctions with the parent dataset, as shown in Table 1. Therefore, when applying our framework to parent models"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.523,
+                0.091,
+                0.872,
+                0.271
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.521,
+                0.288,
+                0.871,
+                0.304
+            ],
+            "angle": 0,
+            "content": "Figure 4: Learning curves of different TL methods."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.329,
+                0.884,
+                0.378
+            ],
+            "angle": 0,
+            "content": "with relatively limited resources, the choice of the number of frozen decoder layers needs to be carefully considered to achieve optimal results."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.387,
+                0.836,
+                0.403
+            ],
+            "angle": 0,
+            "content": "5.2 Effect of Hyper-parameters \\(\\lambda\\) and \\(\\tau\\)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.408,
+                0.884,
+                0.553
+            ],
+            "angle": 0,
+            "content": "Hyper-parameter \\(\\lambda\\) is crucial to controlling the influence of the two losses within the \\(L_{child}\\). In this part, we set \\(\\lambda\\) to \\(\\{2.0, 3.0, 4.0, 5.0\\}\\) to investigate the impact of different values of \\(\\lambda\\) on the performance of the child model. The corresponding SacreBLEU scores are presented in Table 3. For both \\(\\mathrm{Tr} \\rightarrow \\mathrm{En}\\) and \\(\\mathrm{Hu} \\rightarrow \\mathrm{En}\\) translations, the best performances are obtained when \\(\\lambda\\) is set to 4.0. Hence, we set \\(\\lambda\\) as 4.0 for all experiments involving \\(L_{dist}\\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.553,
+                0.884,
+                0.73
+            ],
+            "angle": 0,
+            "content": "In addition, we also conduct experiments with varying values of \\(\\tau\\) during the training process of the child model, while keeping \\(\\lambda\\) fixed at 4.0. As illustrated in Table 3, we can find that the performance of the child model is sensitive to \\(\\tau\\) and the performance is best when \\(\\tau\\) is set to 2.0. We argue that this is because minimizing the KL divergence is difficult, but using a larger \\(\\tau\\) (e.g., 3.0) may diminish the information from the intermediate model, which is not helpful in improving the performance of the child model."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.741,
+                0.675,
+                0.756
+            ],
+            "angle": 0,
+            "content": "5.3 Ablation Study"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.761,
+                0.884,
+                0.922
+            ],
+            "angle": 0,
+            "content": "We conduct an ablation study of the PDF strategy, \\( L_{dist} \\), and Step 2 to explore their effects on our framework. We present the performance of four variants of TSFT as follows: 1) w/o PDF. During the training process of Step 1, we do not freeze any layers of the intermediate model, fine-tuning all parameters in every epoch. 2) w/o \\( L_{dist} \\). In Step 2, we eliminate the distillation loss between the encoders of the intermediate and child models, conducting fine-tuning of the child model using"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3220"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.207,
+                0.094,
+                0.388,
+                0.22
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.258,
+                0.23,
+                0.332,
+                0.243
+            ],
+            "angle": 0,
+            "content": "(a) TM-TL"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.406,
+                0.094,
+                0.59,
+                0.22
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.453,
+                0.23,
+                0.545,
+                0.243
+            ],
+            "angle": 0,
+            "content": "(b) ConsistTL"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.615,
+                0.094,
+                0.792,
+                0.219
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.647,
+                0.23,
+                0.76,
+                0.244
+            ],
+            "angle": 0,
+            "content": "(c) TSFT (Step 1)"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.26,
+                0.882,
+                0.303
+            ],
+            "angle": 0,
+            "content": "Figure 5: Sentence representations after using T-SNE dimensionality reduction. The blue points denote the output from the parent model, and the red points denote the output from the fine-tuned models obtained from different transfer learning methods."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.328,
+                0.49,
+                0.633
+            ],
+            "angle": 0,
+            "content": "\\(L_{ce}\\) exclusively. 3) w/o Step 2. We evaluate the translation performance of the intermediate model. 4) w/o Step \\(2 + \\mathrm{PDF}\\). Based on 3), we do not freeze any layers of the intermediate model during Step 1. We conduct experiments on \\(\\mathrm{Tr} \\rightarrow \\mathrm{En}\\) and \\(\\mathrm{Hu} \\rightarrow \\mathrm{En}\\) translations, which correspondingly represent the largest and smallest datasets among those applied in our main experiments. The results are shown in Table 4. It is evident that excluding the PDF strategy, \\(L_{dist}\\), or Step 2 resulting in a deterioration of the translation quality, underscoring the efficacy of these components within TSFT. The experimental results show that PDF has a greater impact than \\(L_{dist}\\). Further, we observe that PDF can effectively improve the translation performance of the intermediate model and benefit the child model. This observation shows that retaining the performance of the parent model is crucial for improving the performance of the child model."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.644,
+                0.416,
+                0.66
+            ],
+            "angle": 0,
+            "content": "5.4 Comparison of Learning Curves"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.665,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": "A learning curve represents a model's learning performance throughout the duration of training and is a widely employed diagnostic tool in machine learning (Kambhatla et al., 2022; Bao et al., 2023). In this section, we present the validation learning curve to assess the generalization capabilities of TM-TL, ConsistTL, and TSFT by using the Sacre-BLEU score as the criterion. Figure 4 illustrates the learning curves of child models trained with three transfer learning methods. Compared with TM-TL and ConsistTL, TSFT exhibits superior initial performance and convergence speed. Note that the TSFT curve delineates the performance of the model fine-tuned after Step 1. This observation emphasizes the effectiveness of fine-tuning the intermediate model in enhancing the final model's"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.328,
+                0.885,
+                0.522
+            ],
+            "angle": 0,
+            "content": "performance, which can be attributed to the augmentation of adaptability to child data consequent to the fine-tuning process in Step 1. Besides, as the training progresses into the stable phase, we can find that the performance of the child model under the TSFT framework is consistently higher than that of TM-TL and ConsistTL. It is noteworthy that, similar to TM-TL and ConsistTL, TSFT does not utilize additional data or resources. Thus, the performance improvement of the child model can be attributed to the effectiveness of the pre-finetune process."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.535,
+                0.857,
+                0.551
+            ],
+            "angle": 0,
+            "content": "5.5 Sentence Representation Visualization"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.557,
+                0.885,
+                0.848
+            ],
+            "angle": 0,
+            "content": "In our framework, the intermediate model is used to adjust the parent parameters to perform well when using child source sentences as input (Section 3.2). Thus, in this section, we visualize the target-side sentence representations of the De-En parent model and Hu-En models obtained from different transfer learning methods. We utilize the T-SNE method (Hinton and Roweis, 2002) to project the representations into a 2-dimensional space, as shown in Figure 5. This figure shows that TM-TL struggles to align the child representations with the parent representations. ConsistL slightly reduces the discrepancy between the parent and child representations, whereas the intermediate model from TSFT makes the representations much more similar. This observation shows that our fine-tuned intermediate model can produce similar outputs to the parent model even with different source languages."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.862,
+                0.642,
+                0.877
+            ],
+            "angle": 0,
+            "content": "6 Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.891,
+                0.885,
+                0.922
+            ],
+            "angle": 0,
+            "content": "In this paper, we propose TSFT: a novel two-step fine-tuning framework for low-resource NMT."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.519,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3221"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.493,
+                0.295
+            ],
+            "angle": 0,
+            "content": "TSFT incorporates an intermediate (child) model to pre-fine-tune the parent model to fit the child data. The intermediate model is initialized with the parent model and then fine-tuned on the child source data in the first step. We propose freezing partial decoder layers when fine-tuning the intermediate model to alleviate catastrophic forgetting. In the second step, TSFT initializes the child model with the intermediate model and fine-tunes the child model on the parallel data using the cross-entropy and proposed distillation losses. Experimental results on five low-resource translations demonstrate the effectiveness of our proposed TSFT."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.31,
+                0.221,
+                0.325
+            ],
+            "angle": 0,
+            "content": "Limitations"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.339,
+                0.49,
+                0.548
+            ],
+            "angle": 0,
+            "content": "When using our proposed framework, two fine-tuning steps are necessary to obtain the final child model. Therefore, compared to one-step transfer learning methods in NMT, TSFT may require more training time and computation resources to transfer parent knowledge to the child model. Nevertheless, it is important to note that TSFT does not introduce additional time or computing resource consumption during inference. Besides, TSFT is designed for transfer learning scenarios when the target languages of the parent and child models are identical. We will try transferring different target languages in the future."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.564,
+                0.267,
+                0.579
+            ],
+            "angle": 0,
+            "content": "Ethics Statement"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.593,
+                0.489,
+                0.658
+            ],
+            "angle": 0,
+            "content": "This study uses only publicly accessible datasets and models that permit academic research. The preprocessing tools and model training toolkit are open-sourced without copyright conflicts."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.673,
+                0.287,
+                0.689
+            ],
+            "angle": 0,
+            "content": "Acknowledgements"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.702,
+                0.49,
+                0.798
+            ],
+            "angle": 0,
+            "content": "We would like to thank the anonymous reviewers for their helpful comments. This work is supported by the 2020 Catalyst: Strategic New Zealand - Singapore Data Science Research Programme Fund by Ministry of Business, Innovation and Employment (MBIE), New Zealand."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.829,
+                0.214,
+                0.844
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.116,
+                0.854,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Roee Aharoni, Melvin Johnson, and Orhan First. 2019. Massively multilingual neural machine translation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.511,
+                0.086,
+                0.885,
+                0.153
+            ],
+            "angle": 0,
+            "content": "Alham Fikri Aji, Nikolay Bogoychev, Kenneth Heafield, and Rico Sennrich. 2020. In neural machine translation, what does transfer learning transfer? In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.51,
+                0.164,
+                0.885,
+                0.232
+            ],
+            "angle": 0,
+            "content": "Guangsheng Bao, Zhiyang Teng, and Yue Zhang. 2023. Target-side augmentation for document-level machine translation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.51,
+                0.242,
+                0.885,
+                0.295
+            ],
+            "angle": 0,
+            "content": "Alexandre Bérard. 2021. Continual learning in multilingual NMT via language-specific embeddings. In Proceedings of the Sixth Conference on Machine Translation."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.51,
+                0.307,
+                0.884,
+                0.388
+            ],
+            "angle": 0,
+            "content": "Deng Cai, Yan Wang, Huayang Li, Wai Lam, and Lemao Liu. 2021. Neural machine translation with monolingual translation memory. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.51,
+                0.399,
+                0.885,
+                0.466
+            ],
+            "angle": 0,
+            "content": "Chenhui Chu, Raj Dabre, and Sadao Kurohashi. 2017. An empirical comparison of domain adaptation methods for neural machine translation. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.51,
+                0.477,
+                0.884,
+                0.544
+            ],
+            "angle": 0,
+            "content": "Juuso Eronen, Michal Ptaszynski, Karol Nowakowski, Zheng Lin Chia, and Fumito Masui. 2023. Improving polish to english neural machine translation with transfer learning: Effects of data volume and language similarity. arXiv preprint arXiv:2306.00660."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.51,
+                0.555,
+                0.884,
+                0.648
+            ],
+            "angle": 0,
+            "content": "Naman Goyal, Cynthia Gao, Vishrav Chaudhary, Peng-Jen Chen, Guillaume Wenzek, Da Ju, Sanjana Krishnan, Marc'Aurelio Ranzato, Francisco Guzmán, and Angela Fan. 2022. The Flores-101 evaluation benchmark for low-resource and multilingual machine translation. Transactions of the Association for Computational Linguistics."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.51,
+                0.659,
+                0.885,
+                0.714
+            ],
+            "angle": 0,
+            "content": "Shuhao Gu and Yang Feng. 2020. Investigating catastrophic forgetting during continual training for neural machine translation. In Proceedings of the 28th International Conference on Computational Linguistics."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.51,
+                0.724,
+                0.885,
+                0.791
+            ],
+            "angle": 0,
+            "content": "Shuhao Gu, Yang Feng, and Wanying Xie. 2021. Pruning-then-expanding model for domain adaptation of neural machine translation. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.51,
+                0.803,
+                0.884,
+                0.87
+            ],
+            "angle": 0,
+            "content": "Dengji Guo, Zhengrui Ma, Min Zhang, and Yang Feng. 2022. Prediction difference regularization against perturbation for neural machine translation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.51,
+                0.88,
+                0.884,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Geoffrey E Hinton and Sam Roweis. 2002. Stochastic neighbor embedding. Advances in neural information processing systems."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3222"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.152
+            ],
+            "angle": 0,
+            "content": "Aizhan Imankulova, Raj Dabre, Atsushi Fujita, and Kenji Imamura. 2019. Exploiting out-of-domain parallel data through multilingual transfer learning for low-resource neural machine translation. In Proceedings of Machine Translation Summit XVII."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.162,
+                0.489,
+                0.228
+            ],
+            "angle": 0,
+            "content": "Baijun Ji, Zhirui Zhang, Xiangyu Duan, Min Zhang, Boxing Chen, and Weihua Luo. 2020. Cross-lingual pre-training based transfer for zero-shot neural machine translation. In Proceedings of the AAAI conference on artificial intelligence."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.238,
+                0.489,
+                0.304
+            ],
+            "angle": 0,
+            "content": "Nishant Kambhatla, Logan Born, and Anoop Sarkar. 2022. Ciphertext based data augmentation for neural machine translation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.313,
+                0.489,
+                0.379
+            ],
+            "angle": 0,
+            "content": "Huda Khayrallah, Brian Thompson, Matt Post, and Philipp Koehn. 2020. Simulated multiple reference training improves low-resource machine translation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.388,
+                0.489,
+                0.454
+            ],
+            "angle": 0,
+            "content": "Yunsu Kim, Yingbo Gao, and Hermann Ney. 2019a. Effective cross-lingual transfer of neural machine translation models without shared vocabularies. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.464,
+                0.489,
+                0.557
+            ],
+            "angle": 0,
+            "content": "Yunsu Kim, Petre Petrov, Pavel Petrushkov, Shahram Khadivi, and Hermann Ney. 2019b. Pivot-based transfer learning for neural machine translation between non-english languages. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.566,
+                0.489,
+                0.618
+            ],
+            "angle": 0,
+            "content": "Diederik P Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In Proceedings of the 3rd International Conference for Learning Representations."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.628,
+                0.489,
+                0.707
+            ],
+            "angle": 0,
+            "content": "James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. 2017. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.716,
+                0.489,
+                0.77
+            ],
+            "angle": 0,
+            "content": "Tom Kocmi and Ondrej Bojar. 2018. Trivial transfer learning for low-resource neural machine translation. Proceedings of the Third Conference on Machine Translation: Research Papers."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.779,
+                0.489,
+                0.833
+            ],
+            "angle": 0,
+            "content": "Tom Kocmi and Ondrej Bojar. 2020. Efficiently reusing old models across languages via transfer learning. In Proceedings of the 22nd Annual Conference of the European Association for Machine Translation."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.842,
+                0.489,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Surafel M Lakew, Alija Erofeeva, Matteo Negri, Marcello Federico, and Marco Turchi. 2018. Transfer learning in multilingual neural machine translation with dynamic vocabulary. In Proceedings of the 15th International Conference on Spoken Language Translation."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.086,
+                0.883,
+                0.166
+            ],
+            "angle": 0,
+            "content": "Zhaocong Li, Xuebo Liu, Derek F. Wong, Lidia S. Chao, and Min Zhang. 2022. ConsistTL: Modeling consistency in transfer learning for low-resource neural machine translation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.177,
+                0.883,
+                0.217
+            ],
+            "angle": 0,
+            "content": "J. Lin. 1991. Divergence measures based on the shannon entropy. IEEE Transactions on Information Theory."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.229,
+                0.883,
+                0.309
+            ],
+            "angle": 0,
+            "content": "Yu-Hsiang Lin, Chian-Yu Chen, Jean Lee, Zirui Li, Yuyan Zhang, Mengzhou Xia, Shruti Rijhwani, Junxian He, Zhisong Zhang, Xuezhe Ma, et al. 2019. Choosing transfer languages for cross-lingual learning. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.321,
+                0.883,
+                0.4
+            ],
+            "angle": 0,
+            "content": "Shudong Liu, Xuebo Liu, Derek F Wong, Zhaocong Li, Wenxiang Jiao, Lidia S Chao, and Min Zhang. 2023. kNN-TL: k-nearest-neighbor transfer learning for low-resource neural machine translation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.412,
+                0.883,
+                0.464
+            ],
+            "angle": 0,
+            "content": "Gongxu Luo, Yating Yang, Yang Yuan, Zhanheng Chen, and Aizimaiti Ainiwaer. 2019. Hierarchical transfer learning architecture for low-resource neural machine translation. IEEE Access."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.477,
+                0.883,
+                0.543
+            ],
+            "angle": 0,
+            "content": "Mieradilijiang Maimaiti, Yang Liu, Huanbo Luan, and Maosong Sun. 2019. Multi-round transfer learning for low-resource nmt using multiple high-resource languages. ACM Transactions on Asian and Low-Resource Language Information Processing."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.555,
+                0.883,
+                0.609
+            ],
+            "angle": 0,
+            "content": "Paul Michel and Graham Neubig. 2018. Mtnt: A testbed for machine translation of noisy text. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.62,
+                0.883,
+                0.687
+            ],
+            "angle": 0,
+            "content": "Toan Q Nguyen and David Chiang. 2017. Transfer learning across low-resource, related languages for neural machine translation. In Proceedings of the Eighth International Joint Conference on Natural Language Processing."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.698,
+                0.883,
+                0.764
+            ],
+            "angle": 0,
+            "content": "Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. BLEU: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.777,
+                0.883,
+                0.843
+            ],
+            "angle": 0,
+            "content": "Peyman Passban, Qun Lui, and Andy Way. 2017. Translating low-resource languages by vocabulary adaptation from close counterparts. ACM Transactions on Asian and Low-Resource Language Information Processing."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.854,
+                0.883,
+                0.921
+            ],
+            "angle": 0,
+            "content": "Frithjof Petrick, Jan Rosendahl, Christian Herold, and Hermann Ney. 2022. Locality-sensitive hashing for long context neural machine translation. In Proceedings of the 19th International Conference on Spoken Language Translation."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.511,
+                0.086,
+                0.883,
+                0.921
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3223"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.086,
+                0.486,
+                0.125
+            ],
+            "angle": 0,
+            "content": "Matt Post. 2018. A call for clarity in reporting BLEU scores. In Proceedings of the Third Conference on Machine Translation."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.135,
+                0.487,
+                0.188
+            ],
+            "angle": 0,
+            "content": "Keisuke Sakaguchi, Kevin Duh, Matt Post, and Benjamin Van Durme. 2017. Robsut wrod reocginiton via semi-character recurrent neural network. In Thirty-first AAAI conference on artificial intelligence."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.197,
+                0.487,
+                0.262
+            ],
+            "angle": 0,
+            "content": "Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Neural machine translation of rare words with subword units. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.272,
+                0.487,
+                0.31
+            ],
+            "angle": 0,
+            "content": "Hendra Setiawan, Zhongqiang Huang, and Rabih Zbib. 2018. BBN's low-resource machine translation for the LoReHLT 2016 evaluation. Machine Translation."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.32,
+                0.487,
+                0.411
+            ],
+            "angle": 0,
+            "content": "Brian Thompson, Jeremy G Winnup, Huda Khayrallah, Kevin Duh, and Philipp Koehn. 2019. Overcoming catastrophic forgetting during domain adaptation of neural machine translation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.421,
+                0.487,
+                0.471
+            ],
+            "angle": 0,
+            "content": "Jörg Tiedemann. 2012. Parallel data, tools and interfaces in OPUS. In Proceedings of the Eighth International Conference on Language Resources and Evaluation."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.482,
+                0.486,
+                0.547
+            ],
+            "angle": 0,
+            "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.557,
+                0.487,
+                0.622
+            ],
+            "angle": 0,
+            "content": "Yuqiao Wen, Zichao Li, Wenyu Du, and Lili Mou. 2023. f-Divergence minimization for sequence-level knowledge distillation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.632,
+                0.487,
+                0.697
+            ],
+            "angle": 0,
+            "content": "Mengzhou Xia, Xiang Kong, Antonios Anastasopoulos, and Graham Neubig. 2019. Generalized data augmentation for low-resource translation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.706,
+                0.487,
+                0.797
+            ],
+            "angle": 0,
+            "content": "Pengcheng Yin, Hao Fang, Graham Neubig, Adam Pauls, Emmanouil Antonios Platanios, Yu Su, Sam Thomson, and Jacob Andreas. 2021. Compositional generalization for neural semantic parsing via span-level supervised attention. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.807,
+                0.487,
+                0.859
+            ],
+            "angle": 0,
+            "content": "Tianyi Zhang, Varsha Kishore, Felix Wu0, Kilian Q. Weinberger, and Yoav Artzi. 2020. BERTScore: Evaluating text generation with BERT. In International Conference on Learning Representations."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.868,
+                0.487,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Jiawei Zhou and Phillip Keung. 2020. Improving nonautoregressive neural machine translation with monolingual data. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.086,
+                0.487,
+                0.92
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.086,
+                0.881,
+                0.152
+            ],
+            "angle": 0,
+            "content": "Barret Zoph, Deniz Yuret, Jonathan May, and Kevin Knight. 2016. Transfer learning for low-resource neural machine translation. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3224"
+        }
+    ]
+]

2024/A Novel Two-step Fine-tuning Framework for Transfer Learning in Low-Resource Neural Machine Translation/07401a9f-9a22-4566-a6eb-41c799e41ff7_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5f644a9e8f626f529492cfbd37bbd817654108475696be2c89efebb23500ff1f
+size 970013

2024/A Novel Two-step Fine-tuning Framework for Transfer Learning in Low-Resource Neural Machine Translation/full.md

@@ -0,0 +1,290 @@
+# A Novel Two-step Fine-tuning Framework for Transfer Learning in Low-Resource Neural Machine Translation
+
+Yuan Gao, Feng Hou*, Ruili Wang
+
+School of Mathematical and Computational Science, Massey University, New Zealand
+
+{y.gao, f.hou, ruili.wang}@massey.ac.nz
+
+# Abstract
+
+Existing transfer learning methods for neural machine translation typically use a well-trained translation model (i.e., a parent model) of a high-resource language pair to directly initialize a translation model (i.e., a child model) of a low-resource language pair, and the child model is then fine-tuned with corresponding datasets. In this paper, we propose a novel two-step fine-tuning (TSFT) framework for transfer learning in low-resource neural machine translation. In the first step, we adjust the parameters of the parent model to fit the child language by using the child source data. In the second step, we transfer the adjusted parameters to the child model and fine-tune it with a proposed distillation loss for efficient optimization. Our experimental results on five low-resource translations demonstrate that our framework yields significant improvements over various strong transfer learning baselines. Further analysis demonstrated the effectiveness of different components in our framework.
+
+# 1 Introduction
+
+Neural machine translation (NMT) has achieved superior performance in terms of both fluency and adequacy for high-resource languages (Vaswani et al., 2017; Zhou and Keung, 2020; Cai et al., 2021; Guo et al., 2022). With the introduction of the attention mechanism (Yin et al., 2021; Petrick et al., 2022), NMT has been proven to be efficient and powerful in modeling long-distance dependencies. However, the performance of NMT systems deteriorates dramatically when insufficient parallel data are available for training (Sakaguchi et al., 2017; Michel and Neubig, 2018; Aharoni et al., 2019; Goyal et al., 2022). The scarcity of parallel corpora intensely limits the performance of an NMT system on low-resource languages.
+
+Transfer learning is a learning paradigm for addressing the data scarcity problem (Zoph et al.,
+
+![](images/dd040f27afdcbd4a07e347deab3de7619eadc964a52b708a4c36518fb20fb5c1.jpg)  
+Figure 1: Comparison between vanilla transfer learning framework (a) and TSFT (b). Our proposed TSFT incorporates an intermediate model to pre-fine-tune the parent parameters to fit the child data.
+
+2016; Nguyen and Chiang, 2017; Li et al., 2022). For NMT, transfer learning aims to transfer the knowledge from a well-trained high-resource translation model (i.e., a parent model, e.g., English  $\rightarrow$  German) to a low-resource translation model (i.e., a child model, e.g., English  $\rightarrow$  the Māori language). Prior transfer learning methods in NMT (Zoph et al., 2016; Chu et al., 2017) primarily achieve knowledge transfer by initializing the parameters of the child model with the parent model and fine-tuning the child model on the corresponding data. Such direct transfer of knowledge raises a vocabulary mismatch problem (Lakew et al., 2018; Lin et al., 2019; Kocmi and Bojar, 2020), and results in unsatisfied results for low-resource translations. Some methods have been proposed to alleviate the vocabulary mismatch problem, such as constructing joint dictionaries or employing a crosslingual token mapping technique (Passban et al., 2017; Kocmi and Bojar, 2018; Kim et al., 2019a). Additionally, Aji et al. (2020) proposed a token matching method that simply duplicates the embed
+
+dings of overlapping tokens from the parent model to the child model.
+
+Recently, based on the work of Aji et al. (2020), Li et al. (2022) proposed ConsistTL that uses the predictions of the parent model to continuously provide soft targets during the fine-tuning of the child model. However, given the differences between the source inputs of the parent and the child translation tasks, the parent model is not an optimal starting point for the single-step fine-tuning of the child model using limited parallel child data. Therefore, it is necessary to pre-fine-tune the parent model to fit the child language before initializing the child model with it.
+
+Building upon this insight, we propose a simple yet effective transfer learning framework, named Two-Step Fine-Tuning (TSFT), for low-resource NMT. As shown in Figure 1, we introduce an intermediate (child) model initialized with the parent model to adjust the parent parameters to fit the child language. TSFT involves two fine-tuning steps. In the first step, we feed child source sentences (i.e., monolingual data) and meaning-matched sentences in the parent source language into the intermediate and the parent models, respectively. Then, the intermediate model is fine-tuned with the objective of aligning probability distributions from the parent and intermediate models, aiming to adjust the parameters transferred from the parent model to perform well with child source sentences. Additionally, we propose a regularization-based strategy that can improve the translation performance of the intermediate model and benefit the child model. Note that we apply the token matching method to alleviate the vocabulary mismatch problem in the first step. In the second step, we transfer the adjusted parameters from the intermediate model to the child model and fine-tune the entire child model on the pertinent parallel data, employing both a cross-entropy loss and a proposed distillation loss. Extensive experiments on five low-resource translations show that TSFT surpasses the strongest baseline method with up to 1.2 SacreBLEU points. The ablation study demonstrates the effectiveness of different components within TSFT.
+
+Our contributions can be summarized as follows:
+
+- We propose a novel two-step fine-tuning framework for low-resource NMT, which introduces an intermediate (child) model to fit parent parameters for the data of child languages before initializing the child model with
+
+the parent model.
+
+- We propose a regularization-based strategy for fine-tuning the intermediate model and a distillation loss for fine-tuning the child model.  
+- We validate our method by extensive experiments on various low-resource translations and achieve improved performance compared to various transfer learning methods.
+
+# 2 Related work
+
+Existing studies have demonstrated the success of transfer learning for low-resource NMT (Lin et al., 2019; Imankulova et al., 2019; Ji et al., 2020; Eronen et al., 2023). Zoph et al. (2016) first introduced transfer learning into the field of NMT and proposed a parent-child framework, where parameters from a pre-trained parent model are directly transferred to a new child model with a shared target language. Subsequent research largely builds upon the parent-child framework and tends to leverage highly related parent language to perform transfer learning (Passban et al., 2017; Setiawan et al., 2018). However, the languages closely related to low-resource languages are also low-resourced (Nguyen and Chiang, 2017; Xia et al., 2019) and offer only modest performance improvements. Thus, researchers focused on identifying the critical factors for the effectiveness of the parent language. Experimental results from (Lin et al., 2019; Aji et al., 2020) emphasized that linguistic or geographical distance does not appear as important as the size of the parent data (Lin et al., 2019; Aji et al., 2020). This insight expands the range of parent languages available for transfer learning, and alleviates the limitations of highly related parent languages. Consequently, later researchers shifted their attention to parent languages with low relatedness but high-resourced. However, this exacerbates the vocabulary mismatch problem, posing a new challenge to transfer learning.
+
+One solution to the vocabulary mismatch problem is to build a joint dictionary before training a parent model (Kocmi and Bojar, 2018; Kim et al., 2019b). However, this restricts the applicability of a pre-trained parent model to a specific child model only. To overcome this limitation, Kim et al. (2019a) proposed pre-training a language-agnostic cross-lingual word embedding independently from the parent model. Concurrently, token matching methods also show their effectiveness in transfer
+
+learning without requiring additional training efforts (Aji et al., 2020; Kocmi and Bojar, 2020). Some other methods introduce highly related intermediate languages to gradually narrow the vocabulary disparity (Luo et al., 2019; Maimaiti et al., 2019). These methods take advantage of both large-scale data sources and syntactic similarity in the intermediate language.
+
+Recently, Li et al. (2022) incorporated the idea of consistency learning into transfer learning based on the work of Aji et al. (2020) and proposed a novel transfer learning method called ConsistTL. This method enables the child model to utilize the parent model during fine-tuning. Subsequently, Liu et al. (2023) proposed kNN-TL, which extends ConsistTL by integrating a k-nearest neighbor (kNN) module, allowing the child model to utilize the parent model during inference. While our method also builds on ConsistTL, we focus on enhancing the child model's performance during fine-tuning. Thus, our work is orthogonal to kNN-TL.
+
+# 3 Method
+
+In this section, we begin by providing an overview of the basic concepts behind transfer learning and then present our transfer learning framework, TSFT, in detail.
+
+# 3.1 Transfer Learning Primary
+
+Given a source sentence  $x = \{x_{1},\ldots ,x_{I}\}$ , the objective of an NMT model is to translate it to a new sentence  $y = \{y_{1},\dots ,y_{J}\}$  in a target language, where the source sentence and target sentence have lengths  $I$  and  $J$ , respectively. A typical NMT model is composed of an encoder and a decoder. The encoder is designed to extract high-level semantic information from the source sentences and represent them as hidden states  $H_{e}$ . The decoder generates the output probability  $P(y_{i}|H_{e},y_{< i})$  of the next target token  $y_{i}$ . An NMT model is trained on a parallel corpus by minimizing the cross-entropy (CE) loss between the predicted sentence and the ground-truth translation as follows:
+
+$$
+L _ {c e} = - \sum_ {i = 1} ^ {J} \log P \left(y _ {i} \mid y _ {<   i}, x, \theta\right), \tag {1}
+$$
+
+where  $\theta$  is the parameters of the entire NMT model.
+
+Transfer learning has been widely used when only limited training datasets are available for the
+
+problem at hand. It transfers the knowledge acquired from large-scale data to enhance the model performance under low-resource conditions. Transfer learning typically follows a parent-child framework (Zoph et al., 2016), where it involves reusing the parameters  $\theta_{p}$  from a pre-trained parent model to initialize part or all parameters of a child model. In the field of NMT, the parent model  $\mathcal{M}_p$  is initially trained on a high-resourced parallel dataset  $D_{p} = \{X_{p},Y_{p}\}$ , while there is only a limited-sized dataset  $D_{c} = \{X_{c},Y_{c}\}$  available to the child model  $\mathcal{M}_c$ . After the initialization step, the child model can be fine-tuned on  $D_{c}$ , which is also optimized through the minimization of the CE loss.
+
+# 3.2 Two-step Fine-tuning
+
+For NMT, an ideal transfer learning framework should enable the parent model to exert its complete capabilities on the child task. However, owing to the disparities between the parent and child languages, the current one-step fine-tuning transfer learning framework struggles to adjust the parameters of the parent model to fit the child source language under the constraints of limited child data.
+
+The idea of TSFT is simple: before initializing the child model with the parent model, we first adjust the parameters of the parent model to enhance its congruity with the child source language. In this work, we propose to introduce an intermediate model, denoted as  $\mathcal{M}_a$ , to make the parameters of the parent model fit for the child data. Specifically, we initialize the intermediate model with the parent model and pre-fine-tune it by using the source side sentences of the child data, then fine-tune the child model with both the source and target child training data. Therefore, we design TSFT as a two-step framework, as shown in Figure 2.
+
+Step 1: Intermediate Fine-tuning After initializing the intermediate model with a well-trained parent model, we aim to equip the intermediate model with the ability to utilize child source sentences as input for target language generation. Since the intermediate model and the parent model share the same target language, it is crucial to retain the generation ability of the parent model. Therefore, we input the source-side sentences of the child data to the intermediate model and the parent model and utilize the predicted distribution of the parent model as the soft label for fine-tuning.
+
+However, it is infeasible to directly input child
+
+![](images/26b1189479dd93c41352c7e01867b76afb5ca5c53a401e01682bf2ac98306d8f.jpg)  
+Figure 2: Our proposed transfer learning framework TSFT for low-resource NMT. In Step 1, the loss function  $L_{inter}$  is used to optimize the intermediate model. In Step 2, the child model is optimized by  $L_{child}$ . The blue icy blocks are initialized with the parent model and frozen. The input German sentences are produced through back-translation.
+
+source sentences into the parent model, given that the parent and child models have different source languages. Thus, we need a meaning-matched sentence for each child source sentence in the parent source language. In the context of low-resource translations, parallel data for non-English-centric is often limited in size or entirely absent, making it difficult to meet the requirements for intermediate fine-tuning. Therefore, we adopt the method of Li et al. (2022) to generate pseudo parent data  $D_{p*} = \{X_{p*}, Y_c\}$  by using a reversed parent model, where each  $x_{p*} \in X_{p*}$  is aligned with  $y_c \in Y_c$ . Although such a method requires training a reverse parent model, it effectively generates meaning-matched input sentences for the parent model. In addition, we use the following loss function to optimize the intermediate model:
+
+$$
+L _ {i n t e r} = \sum_ {i = 1} ^ {J} F _ {d} [ P _ {i n t e r} (y _ {i}), P _ {p a r e n t} (y _ {i}) ], \quad (2)
+$$
+
+where  $F_{d}$  is a distribution measurement method, in this work, we choose Jensen-Shannon (JS) divergence (Lin, 1991; Wen et al., 2023) as our  $F_{d}$ . Our preliminary experiments find that JS divergence outperforms using Kullback-Leibler (KL) divergence when taking  $P_{inter}(y_i)$  as the first item and  $P_{parent}(y_i)$  as the second one.  $P_{*}(y_i)$  represents the prediction distributions of translation models at time step  $i$ , which is conditioned on the input sentence and the previous tokens:
+
+$$
+P _ {*} (y _ {i}) = P _ {*} (y _ {i} | x, y _ {<   i}). \tag {3}
+$$
+
+Before fine-tuning the intermediate model, we first apply the token matching method (Aji et al., 2020) that duplicates the embeddings of overlapping tokens from the parent and child vocabularies to alleviate the vocabulary mismatch problem.
+
+Step 2: Child Fine-tuning In the second step, we employ the target-side sentences from the child training data as labels to fine-tune the child model with CE loss, following the general process of transfer learning. Since the encoder of the intermediate model has fine-tuned with the child source sentences, we argue that it encompasses valuable information that can facilitate the child model. Therefore, we extract the encoder outputs,  $P_{*}^{e}(\cdot)$ , from both the intermediate and child models and incorporate a distillation loss  $L_{dist}$  as an extra objective to optimize the child model by minimizing the KL divergence between two output representations:
+
+$$
+L _ {d i s t} = - \sum_ {i = 1} ^ {I} P _ {i n t e r} ^ {e} (x _ {i}) \cdot l o g P _ {c h i l d} ^ {e} (x _ {i}), \quad (4)
+$$
+
+$$
+\begin{array}{l} P _ {*} ^ {e} (x _ {i}) = P _ {*} ^ {e} (x _ {i} | x, \tau) \\ = \frac {\exp \left(z _ {i} / \tau\right)}{\sum_ {j \in V} \exp \left(z _ {j} / \tau\right)}, \tag {5} \\ \end{array}
+$$
+
+where  $I$  denotes the sentence length of a child source sentence,  $z$  denotes the logits output of encoders before log softmax is computed,  $V$  represents the vocabulary, and  $\tau$  is a temperate factor used to smooth the prediction distributions. As we
+
+only reuse the output of encoders, the process of encoder distillation does not add any extra parameters to models. The overall loss is obtained by a weighted sum of  $L_{ce}$  and  $L_{dist}$ :
+
+$$
+L _ {\text {c h i l d}} = L _ {c e} + \lambda L _ {\text {d i s t}}, \tag {6}
+$$
+
+where  $\lambda$  is a balancing hyper-parameter.
+
+Partial Decoder Freeze Regularization-based methods are widely used to alleviate the catastrophic forgetting issue (Kirkpatrick et al., 2017; Gu and Feng, 2020; Gu et al., 2021). While updating all parameters typically yields good results on a new domain, the data distribution difference between the old and new domains can engender the issue of catastrophic forgetting, causing the fine-tuned model to abandon linguistic knowledge learned from previous dataset (Thompson et al., 2019; Bérard, 2021). In this work, we are interested in introducing the regularization-based technique during Step 1 to preserve the predictive capabilities of the parent model. We propose a Partial Decoder Freeze (PDF) strategy to freeze the parameters of the last  $l$  decoder layers of the intermediate model and only update the rest parameters. For the selection of parameters  $l$ , we conducted empirical experiments in Section 5.1.
+
+# 4 Experiments
+
+# 4.1 Settings
+
+Datasets We conduct experiments on five low-resource translation tasks, four of which are from the Global Voices datasets (Tiedemann, 2012; Khayrallah et al., 2020): Polish (Pl), Hungarian (Hu), Indonesian (Id), Catalan (Ca) to English (En), where we use the officially provided training sets, validation sets and test sets in our experiments. The other one is the WMT 2017 Turkish (Tr) to En benchmark. We use newstest2016 as the validation set and newstest2017 as the test set. For the parent models training, we use the German-English dataset following the empirical advice of (Aji et al., 2020; Li et al., 2022). We take the WMT 2017 news translation task as our parent dataset containing around 5.8M paired sentences. The detailed statistics of these parallel corpora are presented in Table 1. For fair comparisons, we adopt the same data preprocess techniques as previous research of TL (Li et al., 2022), which only apply normalization and tokenization to
+
+<table><tr><td>Datasets</td><td># Train</td><td># Valid</td><td># Test</td></tr><tr><td>Global Voices PI - En</td><td>39.9K</td><td>2,000</td><td>2,000</td></tr><tr><td>Global Voices Ca -En</td><td>15.2K</td><td>2,000</td><td>2,000</td></tr><tr><td>Global Voices Id - En</td><td>8.4K</td><td>2,000</td><td>2,000</td></tr><tr><td>Global Voices Hu - En</td><td>7.7K</td><td>2,000</td><td>2,000</td></tr><tr><td>WMT 2017 Tr - En</td><td>196.6K</td><td>3,000</td><td>3,007</td></tr><tr><td>WMT 2017 De - En</td><td>5.8M</td><td>3,000</td><td>3,003</td></tr></table>
+
+Table 1: The statistics of parallel corpora.
+
+parallel sentences by using Moses toolkit<sup>1</sup>. Further, we apply Byte Pair Encoding (BPE) (Sennrich et al., 2016) to address the out-of-vocabulary problem and segment words with 16,000 merge operations for Turkish and 8,000 for the rest.
+
+Model Configuration In our experiments, we implement translation models with fairseq toolkit. We choose the Transformer (Vaswani et al., 2017) as the backbone to implement our framework. We use Transformer_base that consists of 6 encoder and decoder layers with 8 attention heads. The number of dimensions of all sub-layers in the model is set to 512, and the inner layers of feed-forward layers have 2048 dimensions. Our models are trained on 2 Nvidia A100 GPUs. We train our models using Adam (Kingma and Ba, 2015) with  $(\beta_{1},\beta_{2}) = (0.9,0.98)$  and use cross-entropy as criterion with label smoothing  $= 0.1$ . In addition, we train the forward and backward parent model (i.e.,  $\mathrm{De}\rightarrow \mathrm{En}$  and  $\mathrm{En}\rightarrow \mathrm{De}$ ) with the initial learning rate  $1e^{-7}$  and gradually increase till  $1e^{-3}$  within 10,000 warm-up updates. For the models with transfer learning, we set the initial learning rate to  $1e^{-7}$ , and the peak learning rate is  $2e^{-4}$  within 1,000 warm-up steps. Dropout is applied to the output of each sub-layer with a rate of 0.3 to avoid over-fitting. Besides, attention and activation dropouts are also used with a rate of 0.1 and 0.1. We train all models with a maximum of 200 epochs and select the checkpoints with the best BLEU score on the validation set as our final model, where beam search is applied with beam size 5, and the length penalty is 1.
+
+Baselines We use the following baselines to validate our method:
+
+<table><tr><td rowspan="2">Model</td><td colspan="2">Tr→En</td><td colspan="2">Hu→En</td><td colspan="2">Id→En</td><td colspan="2">Ca→En</td><td colspan="2">Pl→En</td></tr><tr><td>BLEU</td><td>BS</td><td>BLEU</td><td>BS</td><td>BLEU</td><td>BS</td><td>BLEU</td><td>BS</td><td>BLEU</td><td>BS</td></tr><tr><td>Vanilla</td><td>17.8</td><td>51.8</td><td>0.9</td><td>0.9</td><td>1.1</td><td>13.2</td><td>1.1</td><td>15.5</td><td>1.5</td><td>18.9</td></tr><tr><td>TL</td><td>17.6</td><td>51.9</td><td>5.9</td><td>27.4</td><td>13.5</td><td>37.7</td><td>21.6</td><td>51.8</td><td>19.9</td><td>55.3</td></tr><tr><td>TM-TL</td><td>18.6</td><td>53.9</td><td>10.6</td><td>41.2</td><td>18.6</td><td>49.9</td><td>25.3</td><td>58.9</td><td>21.4</td><td>58.2</td></tr><tr><td>ConsistTL</td><td>19.3</td><td>55.9</td><td>11.9</td><td>43.9</td><td>19.7</td><td>52.2</td><td>26.6</td><td>60.0</td><td>22.4</td><td>59.9</td></tr><tr><td>TSFT (ours)</td><td>20.0</td><td>56.7</td><td>13.1</td><td>44.6</td><td>20.5</td><td>53.3</td><td>27.7</td><td>60.7</td><td>23.3</td><td>60.5</td></tr></table>
+
+- Vanilla NMT (Vaswani et al., 2017): A bilingual NMT model with Transformer architecture directly trained on low-resource child training data from scratch.  
+- TL (Zoph et al., 2016): The first transfer learning work for NMT, initializing the child model with a parent model except for the source word embeddings. Note that the original work employed a two-layer encoder-decoder LSTM model, whereas we replicate TL using Transformer.  
+- TM-TL (Aji et al., 2020): To transfer embeddings across languages with distinct linguistic characteristics, Token Matching (TM) is proposed to assign the child word embeddings with the same tokens in the parent embeddings. The remaining unmatched tokens are assigned random embeddings as TL.  
+- ConsistTL (Li et al., 2022): Based on TM-TL, ConsistTL is proposed to enhance the child model by incorporating the prediction of the parent model during the fine-tuning of the child model.
+
+Metrics To validate the effectiveness of our proposed framework, we use the following two metrics:
+
+- BLEU (Papineni et al., 2002): Considering the discrepancy among different tokenization processes, we apply the SacreBLEU score (Post, 2018)<sup>3</sup> for all experiments.
+
+Table 2: The SacreBLEU and BERTScore scores of baselines and ours on various translations. "BS" represents BERTScore. Blod indicates the best result. BLEU score reflects that TSFT is significantly better than ConsistTL with t-test  $p < 0.05$ . The number of bootstrap resamples is set to 1,000 to measure the significant difference between results.  
+
+<table><tr><td>Hyper-parameter</td><td>Tr→En</td><td>Hu→En</td></tr><tr><td>(λ = 2.0, τ = 2.0)</td><td>19.9</td><td>13.0</td></tr><tr><td>(λ = 3.0, τ = 2.0)</td><td>19.8</td><td>12.8</td></tr><tr><td>(λ = 4.0, τ = 2.0)</td><td>20.0</td><td>13.1</td></tr><tr><td>(λ = 5.0, τ = 2.0)</td><td>19.9</td><td>12.9</td></tr><tr><td>(λ = 4.0, τ = 0.5)</td><td>19.7</td><td>12.9</td></tr><tr><td>(λ = 4.0, τ = 1.0)</td><td>19.7</td><td>13.1</td></tr><tr><td>(λ = 4.0, τ = 3.0)</td><td>19.4</td><td>13.0</td></tr></table>
+
+Table 3: The SacreBLEU scores on the test set of the  $\mathrm{Tr} \rightarrow \mathrm{En}$  and  $\mathrm{Hu} \rightarrow \mathrm{En}$  translations with different  $\lambda$  and  $\tau$ .
+
+- BERTScore (Zhang et al., 2020): Leveraging a pre-trained BERT model to evaluate the semantic correctness between the predictions and references by cosine similarity.
+
+# 4.2 Main Results
+
+The results on five low-resource translation benchmarks are presented in Table 2. In our experiments, we utilize German as the parent language, and the parent models are pre-trained on a German-to-English dataset. As we can see, our method significantly outperforms the vanilla NMT in terms of both SacreBLEU and BERTScore. Compared with TL and TM-TL, TSFT still achieves significant improvements on all translations. Moreover, our proposed TSFT also has demonstrated superior performance compared to the strongest baseline ConsistTL with up to +1.2 SacreBLEU points and +1.1 BERTScore points. Overall, these results prove that our proposed transfer learning framework TSFT can effectively improve the performance of the child model on low-resource translation tasks.
+
+![](images/8651d18bfa80a35e1e777c1dab62df2914a824210edfdb0a544b9784964a2b44.jpg)  
+Figure 3: The SacreBLEU scores of TSFT with different hyper-parameter  $l$  on  $\mathrm{Tr} \rightarrow \mathrm{En}$  and  $\mathrm{Hu} \rightarrow \mathrm{En}$ .  $\mathrm{De} \Rightarrow \mathrm{Tr} / \mathrm{Hu}$  indicates De is the parent language and  $\mathrm{Tr} / \mathrm{Hu}$  is the child language.
+
+<table><tr><td>Models</td><td>Tr→En</td><td>Hu→En</td></tr><tr><td>TSFT</td><td>20.0</td><td>13.1</td></tr><tr><td>w/o PDF</td><td>19.5</td><td>12.5</td></tr><tr><td>w/o Ldist</td><td>19.8</td><td>12.8</td></tr><tr><td>w/o Step 2</td><td>18.9</td><td>11.2</td></tr><tr><td>w/o Step 2 + PDF</td><td>18.6</td><td>10.6</td></tr></table>
+
+Table 4: The SacreBLEU scores on the test set of the  $\mathrm{Tr}\rightarrow \mathrm{En}$  and  $\mathrm{Hu}\rightarrow \mathrm{En}$  translations with PDF,  $L_{dist}$ , and Step 2 ablation.
+
+# 5 Analysis
+
+# 5.1 Effect of the Number of Freezing Layers
+
+In Section 3.2, we utilize the PDF strategy in Step 1. However, we do not clearly know the optimal number of freezing layers  $l$  that can benefit the child model most. Different numbers of freezing layers would significantly impact the child model performance. Hence, in this section, we conduct a comparative analysis of the impact of different  $l$  on the translation performance of the child model.
+
+Concretely, we still use the  $\mathrm{De}\rightarrow \mathrm{En}$  model as the parent model and select  $\mathrm{Tr}\to \mathrm{En}$  and  $\mathrm{Hu}\to \mathrm{En}$  translations as child tasks. We tune the hyperparameter  $l$  by performing a grid search on  $l\in$ $\{1,2,3,4,5,6\}$ . Figure 3 illustrates the model performance with different values of  $l$ . We can find that the final child models achieve the best performance in  $\mathrm{Tr}\to \mathrm{En}$  and  $\mathrm{Hu}\to \mathrm{En}$  when  $l$  is 5 and 4, respectively. Consequently, we set  $l$  as 5 for  $\mathrm{Tr}\to \mathrm{En}$  translation and 4 for the rest.
+
+Despite a substantial size difference between the  $\mathrm{Tr} \rightarrow \mathrm{En}$  and  $\mathrm{Hu} \rightarrow \mathrm{En}$  datasets, there is not much difference in the choice of the number of layers to freeze. For this phenomenon, we speculate that the distinction between these two child datasets is negligible compared to the size distinctions with the parent dataset, as shown in Table 1. Therefore, when applying our framework to parent models
+
+![](images/4d7af7a37c77509d54c570ab7bd478f725e3e82560ba64a4c6f4c513e8506378.jpg)  
+Figure 4: Learning curves of different TL methods.
+
+with relatively limited resources, the choice of the number of frozen decoder layers needs to be carefully considered to achieve optimal results.
+
+# 5.2 Effect of Hyper-parameters  $\lambda$  and  $\tau$
+
+Hyper-parameter  $\lambda$  is crucial to controlling the influence of the two losses within the  $L_{child}$ . In this part, we set  $\lambda$  to  $\{2.0, 3.0, 4.0, 5.0\}$  to investigate the impact of different values of  $\lambda$  on the performance of the child model. The corresponding SacreBLEU scores are presented in Table 3. For both  $\mathrm{Tr} \rightarrow \mathrm{En}$  and  $\mathrm{Hu} \rightarrow \mathrm{En}$  translations, the best performances are obtained when  $\lambda$  is set to 4.0. Hence, we set  $\lambda$  as 4.0 for all experiments involving  $L_{dist}$ .
+
+In addition, we also conduct experiments with varying values of  $\tau$  during the training process of the child model, while keeping  $\lambda$  fixed at 4.0. As illustrated in Table 3, we can find that the performance of the child model is sensitive to  $\tau$  and the performance is best when  $\tau$  is set to 2.0. We argue that this is because minimizing the KL divergence is difficult, but using a larger  $\tau$  (e.g., 3.0) may diminish the information from the intermediate model, which is not helpful in improving the performance of the child model.
+
+# 5.3 Ablation Study
+
+We conduct an ablation study of the PDF strategy,  $L_{dist}$ , and Step 2 to explore their effects on our framework. We present the performance of four variants of TSFT as follows: 1) w/o PDF. During the training process of Step 1, we do not freeze any layers of the intermediate model, fine-tuning all parameters in every epoch. 2) w/o  $L_{dist}$ . In Step 2, we eliminate the distillation loss between the encoders of the intermediate and child models, conducting fine-tuning of the child model using
+
+![](images/5953f927bc5a3643ef28211f92a6b3339d9850b87ce090a6e3c24cf2a9d77c83.jpg)  
+(a) TM-TL
+
+![](images/408affc16d105e3d1d6d5ad052c8b3fe4390fb1498d545a79a25756a1d67f8dd.jpg)  
+(b) ConsistTL  
+Figure 5: Sentence representations after using T-SNE dimensionality reduction. The blue points denote the output from the parent model, and the red points denote the output from the fine-tuned models obtained from different transfer learning methods.
+
+![](images/9f6ab930631684a907d33bea6ddcc6c2bf444f77e8e41fc23e9c4e8c4a2bfebf.jpg)  
+(c) TSFT (Step 1)
+
+$L_{ce}$  exclusively. 3) w/o Step 2. We evaluate the translation performance of the intermediate model. 4) w/o Step  $2 + \mathrm{PDF}$ . Based on 3), we do not freeze any layers of the intermediate model during Step 1. We conduct experiments on  $\mathrm{Tr} \rightarrow \mathrm{En}$  and  $\mathrm{Hu} \rightarrow \mathrm{En}$  translations, which correspondingly represent the largest and smallest datasets among those applied in our main experiments. The results are shown in Table 4. It is evident that excluding the PDF strategy,  $L_{dist}$ , or Step 2 resulting in a deterioration of the translation quality, underscoring the efficacy of these components within TSFT. The experimental results show that PDF has a greater impact than  $L_{dist}$ . Further, we observe that PDF can effectively improve the translation performance of the intermediate model and benefit the child model. This observation shows that retaining the performance of the parent model is crucial for improving the performance of the child model.
+
+# 5.4 Comparison of Learning Curves
+
+A learning curve represents a model's learning performance throughout the duration of training and is a widely employed diagnostic tool in machine learning (Kambhatla et al., 2022; Bao et al., 2023). In this section, we present the validation learning curve to assess the generalization capabilities of TM-TL, ConsistTL, and TSFT by using the Sacre-BLEU score as the criterion. Figure 4 illustrates the learning curves of child models trained with three transfer learning methods. Compared with TM-TL and ConsistTL, TSFT exhibits superior initial performance and convergence speed. Note that the TSFT curve delineates the performance of the model fine-tuned after Step 1. This observation emphasizes the effectiveness of fine-tuning the intermediate model in enhancing the final model's
+
+performance, which can be attributed to the augmentation of adaptability to child data consequent to the fine-tuning process in Step 1. Besides, as the training progresses into the stable phase, we can find that the performance of the child model under the TSFT framework is consistently higher than that of TM-TL and ConsistTL. It is noteworthy that, similar to TM-TL and ConsistTL, TSFT does not utilize additional data or resources. Thus, the performance improvement of the child model can be attributed to the effectiveness of the pre-finetune process.
+
+# 5.5 Sentence Representation Visualization
+
+In our framework, the intermediate model is used to adjust the parent parameters to perform well when using child source sentences as input (Section 3.2). Thus, in this section, we visualize the target-side sentence representations of the De-En parent model and Hu-En models obtained from different transfer learning methods. We utilize the T-SNE method (Hinton and Roweis, 2002) to project the representations into a 2-dimensional space, as shown in Figure 5. This figure shows that TM-TL struggles to align the child representations with the parent representations. ConsistL slightly reduces the discrepancy between the parent and child representations, whereas the intermediate model from TSFT makes the representations much more similar. This observation shows that our fine-tuned intermediate model can produce similar outputs to the parent model even with different source languages.
+
+# 6 Conclusion
+
+In this paper, we propose TSFT: a novel two-step fine-tuning framework for low-resource NMT.
+
+TSFT incorporates an intermediate (child) model to pre-fine-tune the parent model to fit the child data. The intermediate model is initialized with the parent model and then fine-tuned on the child source data in the first step. We propose freezing partial decoder layers when fine-tuning the intermediate model to alleviate catastrophic forgetting. In the second step, TSFT initializes the child model with the intermediate model and fine-tunes the child model on the parallel data using the cross-entropy and proposed distillation losses. Experimental results on five low-resource translations demonstrate the effectiveness of our proposed TSFT.
+
+# Limitations
+
+When using our proposed framework, two fine-tuning steps are necessary to obtain the final child model. Therefore, compared to one-step transfer learning methods in NMT, TSFT may require more training time and computation resources to transfer parent knowledge to the child model. Nevertheless, it is important to note that TSFT does not introduce additional time or computing resource consumption during inference. Besides, TSFT is designed for transfer learning scenarios when the target languages of the parent and child models are identical. We will try transferring different target languages in the future.
+
+# Ethics Statement
+
+This study uses only publicly accessible datasets and models that permit academic research. The preprocessing tools and model training toolkit are open-sourced without copyright conflicts.
+
+# Acknowledgements
+
+We would like to thank the anonymous reviewers for their helpful comments. This work is supported by the 2020 Catalyst: Strategic New Zealand - Singapore Data Science Research Programme Fund by Ministry of Business, Innovation and Employment (MBIE), New Zealand.
+
+# References
+
+Roee Aharoni, Melvin Johnson, and Orhan First. 2019. Massively multilingual neural machine translation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics.
+
+Alham Fikri Aji, Nikolay Bogoychev, Kenneth Heafield, and Rico Sennrich. 2020. In neural machine translation, what does transfer learning transfer? In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics.
+
+Guangsheng Bao, Zhiyang Teng, and Yue Zhang. 2023. Target-side augmentation for document-level machine translation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics.
+
+Alexandre Bérard. 2021. Continual learning in multilingual NMT via language-specific embeddings. In Proceedings of the Sixth Conference on Machine Translation.
+
+Deng Cai, Yan Wang, Huayang Li, Wai Lam, and Lemao Liu. 2021. Neural machine translation with monolingual translation memory. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing.
+
+Chenhui Chu, Raj Dabre, and Sadao Kurohashi. 2017. An empirical comparison of domain adaptation methods for neural machine translation. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics.
+
+Juuso Eronen, Michal Ptaszynski, Karol Nowakowski, Zheng Lin Chia, and Fumito Masui. 2023. Improving polish to english neural machine translation with transfer learning: Effects of data volume and language similarity. arXiv preprint arXiv:2306.00660.
+
+Naman Goyal, Cynthia Gao, Vishrav Chaudhary, Peng-Jen Chen, Guillaume Wenzek, Da Ju, Sanjana Krishnan, Marc'Aurelio Ranzato, Francisco Guzmán, and Angela Fan. 2022. The Flores-101 evaluation benchmark for low-resource and multilingual machine translation. Transactions of the Association for Computational Linguistics.
+
+Shuhao Gu and Yang Feng. 2020. Investigating catastrophic forgetting during continual training for neural machine translation. In Proceedings of the 28th International Conference on Computational Linguistics.
+
+Shuhao Gu, Yang Feng, and Wanying Xie. 2021. Pruning-then-expanding model for domain adaptation of neural machine translation. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics.
+
+Dengji Guo, Zhengrui Ma, Min Zhang, and Yang Feng. 2022. Prediction difference regularization against perturbation for neural machine translation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics.
+
+Geoffrey E Hinton and Sam Roweis. 2002. Stochastic neighbor embedding. Advances in neural information processing systems.
+
+Aizhan Imankulova, Raj Dabre, Atsushi Fujita, and Kenji Imamura. 2019. Exploiting out-of-domain parallel data through multilingual transfer learning for low-resource neural machine translation. In Proceedings of Machine Translation Summit XVII.  
+Baijun Ji, Zhirui Zhang, Xiangyu Duan, Min Zhang, Boxing Chen, and Weihua Luo. 2020. Cross-lingual pre-training based transfer for zero-shot neural machine translation. In Proceedings of the AAAI conference on artificial intelligence.  
+Nishant Kambhatla, Logan Born, and Anoop Sarkar. 2022. Ciphertext based data augmentation for neural machine translation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics.  
+Huda Khayrallah, Brian Thompson, Matt Post, and Philipp Koehn. 2020. Simulated multiple reference training improves low-resource machine translation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing.  
+Yunsu Kim, Yingbo Gao, and Hermann Ney. 2019a. Effective cross-lingual transfer of neural machine translation models without shared vocabularies. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics.  
+Yunsu Kim, Petre Petrov, Pavel Petrushkov, Shahram Khadivi, and Hermann Ney. 2019b. Pivot-based transfer learning for neural machine translation between non-english languages. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing.  
+Diederik P Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In Proceedings of the 3rd International Conference for Learning Representations.  
+James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. 2017. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences.  
+Tom Kocmi and Ondrej Bojar. 2018. Trivial transfer learning for low-resource neural machine translation. Proceedings of the Third Conference on Machine Translation: Research Papers.  
+Tom Kocmi and Ondrej Bojar. 2020. Efficiently reusing old models across languages via transfer learning. In Proceedings of the 22nd Annual Conference of the European Association for Machine Translation.  
+Surafel M Lakew, Alija Erofeeva, Matteo Negri, Marcello Federico, and Marco Turchi. 2018. Transfer learning in multilingual neural machine translation with dynamic vocabulary. In Proceedings of the 15th International Conference on Spoken Language Translation.
+
+Zhaocong Li, Xuebo Liu, Derek F. Wong, Lidia S. Chao, and Min Zhang. 2022. ConsistTL: Modeling consistency in transfer learning for low-resource neural machine translation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing.  
+J. Lin. 1991. Divergence measures based on the shannon entropy. IEEE Transactions on Information Theory.  
+Yu-Hsiang Lin, Chian-Yu Chen, Jean Lee, Zirui Li, Yuyan Zhang, Mengzhou Xia, Shruti Rijhwani, Junxian He, Zhisong Zhang, Xuezhe Ma, et al. 2019. Choosing transfer languages for cross-lingual learning. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics.  
+Shudong Liu, Xuebo Liu, Derek F Wong, Zhaocong Li, Wenxiang Jiao, Lidia S Chao, and Min Zhang. 2023. kNN-TL: k-nearest-neighbor transfer learning for low-resource neural machine translation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics.  
+Gongxu Luo, Yating Yang, Yang Yuan, Zhanheng Chen, and Aizimaiti Ainiwaer. 2019. Hierarchical transfer learning architecture for low-resource neural machine translation. IEEE Access.  
+Mieradilijiang Maimaiti, Yang Liu, Huanbo Luan, and Maosong Sun. 2019. Multi-round transfer learning for low-resource nmt using multiple high-resource languages. ACM Transactions on Asian and Low-Resource Language Information Processing.  
+Paul Michel and Graham Neubig. 2018. Mtnt: A testbed for machine translation of noisy text. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing.  
+Toan Q Nguyen and David Chiang. 2017. Transfer learning across low-resource, related languages for neural machine translation. In Proceedings of the Eighth International Joint Conference on Natural Language Processing.  
+Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. BLEU: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics.  
+Peyman Passban, Qun Lui, and Andy Way. 2017. Translating low-resource languages by vocabulary adaptation from close counterparts. ACM Transactions on Asian and Low-Resource Language Information Processing.  
+Frithjof Petrick, Jan Rosendahl, Christian Herold, and Hermann Ney. 2022. Locality-sensitive hashing for long context neural machine translation. In Proceedings of the 19th International Conference on Spoken Language Translation.
+
+Matt Post. 2018. A call for clarity in reporting BLEU scores. In Proceedings of the Third Conference on Machine Translation.  
+Keisuke Sakaguchi, Kevin Duh, Matt Post, and Benjamin Van Durme. 2017. Robsut wrod reocginiton via semi-character recurrent neural network. In Thirty-first AAAI conference on artificial intelligence.  
+Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Neural machine translation of rare words with subword units. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics.  
+Hendra Setiawan, Zhongqiang Huang, and Rabih Zbib. 2018. BBN's low-resource machine translation for the LoReHLT 2016 evaluation. Machine Translation.  
+Brian Thompson, Jeremy G Winnup, Huda Khayrallah, Kevin Duh, and Philipp Koehn. 2019. Overcoming catastrophic forgetting during domain adaptation of neural machine translation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies.  
+Jörg Tiedemann. 2012. Parallel data, tools and interfaces in OPUS. In Proceedings of the Eighth International Conference on Language Resources and Evaluation.  
+Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems.  
+Yuqiao Wen, Zichao Li, Wenyu Du, and Lili Mou. 2023. f-Divergence minimization for sequence-level knowledge distillation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics.  
+Mengzhou Xia, Xiang Kong, Antonios Anastasopoulos, and Graham Neubig. 2019. Generalized data augmentation for low-resource translation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics.  
+Pengcheng Yin, Hao Fang, Graham Neubig, Adam Pauls, Emmanouil Antonios Platanios, Yu Su, Sam Thomson, and Jacob Andreas. 2021. Compositional generalization for neural semantic parsing via span-level supervised attention. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics.  
+Tianyi Zhang, Varsha Kishore, Felix Wu0, Kilian Q. Weinberger, and Yoav Artzi. 2020. BERTScore: Evaluating text generation with BERT. In International Conference on Learning Representations.  
+Jiawei Zhou and Phillip Keung. 2020. Improving nonautoregressive neural machine translation with monolingual data. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics.
+
+Barret Zoph, Deniz Yuret, Jonathan May, and Kevin Knight. 2016. Transfer learning for low-resource neural machine translation. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing.

2024/A Novel Two-step Fine-tuning Framework for Transfer Learning in Low-Resource Neural Machine Translation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ad4781624a5ebc0b94dde6d727131e3a463a994ddbaa0df61ae02a1d327b13c6
+size 374496

2024/A Robust Semantics-based Watermark for Large Language Model against Paraphrasing/b1276659-b3bd-43b9-95dd-a7cb02ea87a5_content_list.json

@@ -0,0 +1,1485 @@
+[
+    {
+        "type": "text",
+        "text": "A Robust Semantics-based Watermark for Large Language Models against Paraphrasing",
+        "text_level": 1,
+        "bbox": [
+            149,
+            89,
+            847,
+            130
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Jie Ren $^{1}$ , Han Xu $^{1}$ , Yiding Liu $^{2}$ , Yingqian Cui $^{1}$ , Shuaiqiang Wang $^{2}$ , Dawei Yin $^{2}$ , Jiliang Tang $^{1}$",
+        "bbox": [
+            211,
+            145,
+            789,
+            180
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "<sup>1</sup>Michigan State University, <sup>2</sup>Baidu Inc.",
+        "bbox": [
+            341,
+            180,
+            660,
+            196
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{renjie3, xuhan1, cuiyingq, tangjili}@msu.edu",
+        "bbox": [
+            272,
+            198,
+            727,
+            212
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "liuyiding.tanh@gmail.com, shqiang.wang@gmail.com, yindawei@acm.org",
+        "bbox": [
+            166,
+            214,
+            833,
+            229
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            261,
+            339,
+            275
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Large language models (LLMs) have show their remarkable ability in various natural language tasks. However, there are concerns that LLMs are possible to be used improperly or even illegally. To prevent the malicious usage of LLMs, detecting LLM-generated text becomes crucial in the deployment of LLM applications. Watermarking is an effective strategy to detect the LLM-generated content by encoding a pre-defined secret watermark to facilitate the detection process. However, the majority of existing watermark methods leverage the simple hashes of precedent tokens to partition vocabulary. Such watermarks can be easily eliminated by paraphrase and, correspondingly, the detection effectiveness will be greatly compromised. Thus, to enhance the robustness against paraphrase, we propose a semantics-based watermark framework, SemaMark. It leverages the semantics as an alternative to simple hashes of tokens since the semantic meaning of the sentences will be likely preserved under paraphrase and the watermark can remain robust. Comprehensive experiments are conducted to demonstrate the effectiveness and robustness of SemaMark under different paraphrases. Our code is available at github.com/renjie3/SemaMark.",
+        "bbox": [
+            144,
+            282,
+            460,
+            680
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            687,
+            258,
+            702
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Large language models (LLMs) have shown their great ability in various natural language processing (NLP) tasks like Question Answering (QA) (Lu et al., 2022), reasoning tasks (Wei et al., 2022; Creswell et al., 2022) and code development (Xu et al., 2022). However, tremendous concerns have been raised that LLMs are possible to be used improperly and illegally. For example, indistinguishable fake news are easy to be fabricated (Kreps et al., 2022; Zellers et al., 2019) by language models, which, when disseminated, could instigate widespread panic. Similarly, in the commercial sphere, convincingly generated reviews",
+        "bbox": [
+            112,
+            712,
+            489,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "can manipulate consumer perceptions, leading to unethical business competition (Salminen et al., 2022). Therefore, detecting LLM-generated text has become crucial in the real-world applications of LLMs (Wu et al., 2023; Sadasivan et al., 2023; Xu et al., 2023).",
+        "bbox": [
+            507,
+            261,
+            884,
+            356
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Among diverse methods to detect LLM-generated texts, the watermark strategies have demonstrated outstanding precision (Liu et al., 2023b; Tang et al., 2023; Ren et al., 2024). It is proposed to encode a secret watermark into the generated texts, such that we can tell whether a text is generated by detecting this watermark. One representative strategy (Kirchenbauer et al., 2023a; Yoo et al., 2023) is to encode the watermark based on the \"partition of vocabulary\". In detail, given a language model, these methods devise a mapping from precedent tokens to a particular partition of the vocabulary by a partition function for the consequent token. The partition function leverages the hashes of the input as the seed of a random generator to split the vocabulary to a green list and a red list. During the text generation phase, the consequent token has an increased probability to be sampled from the green list. In this way, the watermark is encoded through the matching between the precedent tokens and the vocabulary partition for the consequent token. The detection is also facilitated by detecting this matching in generated contents. However, recent works (Krishna et al., 2023; Kaddour et al., 2023) reveal that this watermark may be easily eliminated by sentence paraphrasing. Individuals seeking to improperly utilize LLMs without being detected can paraphrase the generated contents, like altering the order and the choices of the words, and only retain the general meaning of the text to achieve their malicious goals like faking news. These paraphrases will change the seed of the partition function, i.e. the token hashes, and as we show in the Section 4.4, the partition function is sensitive to small changes. Consequently,",
+        "bbox": [
+            507,
+            357,
+            884,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "613",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: NAACL 2024, pages 613-625 June 16-21, 2024 ©2024 Association for Computational Linguistics",
+        "bbox": [
+            226,
+            945,
+            769,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "the matching between the precedent tokens and the green list will be disrupted, and the detection effectiveness of the watermark can be dramatically compromised.",
+        "bbox": [
+            112,
+            84,
+            487,
+            148
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we propose to leverage the semantic meaning of precedent token sequences as the seed for partition function, instead of simple hashes of precedent tokens, since the core semantic meaning is expected to be maintained after paraphrase. To achieve this goal, one key obstacle is how to capture the semantics when applying them for the partition function to watermark the generated texts. It is a common practice to quantify the semantics via embeddings (Reimers and Gurevych, 2019; Gao et al., 2021; Li et al., 2020; Giorgi et al., 2021). Embeddings indeed can represent consistent semantics after paraphrase. Since the embeddings are high-dimensional vectors in the continuous space, they often present some minor changes after paraphrase. Although the main semantics are preserved, these minor changes can lead to a substantial difference in the partition of vocabulary because the random generator in the partition function is sensitive to the change of the seed, as shown in Section 4.4.",
+        "bbox": [
+            115,
+            149,
+            489,
+            469
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To overcome the above challenge, i.e., to make the quantified semantics invariant and make the watermark robust under paraphrase, we propose a new watermark method, SemaMark, which discretizes the continuous embedding space. Intuitively, the discretization can coarsen the representation of the embeddings which could tolerate the potential minor changes caused by paraphrase. By proper discretization, the paraphrased semantics could stay in the same discrete section with a high probability and the discretized quantified semantics will likely remain the same even after paraphrase. Therefore, the partition results will not change. However, directly converting the high-dimensional embedding space into discrete is intricate and challenging. For example, discretizing each dimension will lead to a large amount of discrete values which is exponential to the number of dimensions. Thus, the minor changes by paraphrase can still cause the change of discrete values because the number of discrete values are too dense and each discrete value can tolerate only small changes. Therefore, the minor changes of high-dimensional embeddings can have a strong impact on the partition function. To address this problem, SemaMark first uses a MultiLayer Perception (MLP) to condense the continuous high-dimensional embeddings into normalized vectors in 2D space. The vectors are located",
+        "bbox": [
+            115,
+            470,
+            489,
+            920
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "on a unit circle named Normalized Embedding Ring (NE-Ring). Then the condensed NE-Ring is equally divided into various sections, transforming the continuous space into distinct discrete values, i.e., \"semantic values\". Based on the discretization, SemaMark further introduces two strategies to advance the watermark's concealment and to improve the robustness under paraphrase. First, SemaMark leverages the uniformity (Wang and Isola, 2020) of Contrastive Learning(CL) (Chen et al., 2020) to strength the MLP and mitigate the problem that the semantics are unevenly concentrating on some discrete sections on NE-Ring. The unevenly distribution will cause the final discrete semantic values overly monotonous. It raises the concern that the watermark might be cracked by counting token frequency (Zhao et al., 2023). Second, SemaMark utilizes an offset detection method to further enhance the robustness at the boundary of different discrete sections whose semantic values are possibly vulnerable to paraphrase. Comprehensive experiments are conducted to demonstrate the effectiveness and robustness of SemaMark under different paraphrases.",
+        "bbox": [
+            507,
+            84,
+            884,
+            470
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related works",
+        "text_level": 1,
+        "bbox": [
+            509,
+            483,
+            670,
+            497
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "LLM-generated detection. As the development of LLMs, various LLM-generated detection tools have also been proposed. Learning-based methods train a classification model to detect the difference between human-written text and machine-generated text like Guo et al. (2023); Wang et al. (2023); Li et al. (2023). Other works do not rely on the classification model, but try to use the property of the LLM to test whether a given text is generated by LLMs. For example, DetectGPT (Mitchell et al., 2023) assumes that the generated text will have high likelihood. GPT-who (Venkatraman et al., 2023) uses UID-based features to model the unique statistical signature of each LLM and human author for accurate authorship attribution. These methods do not interact the generation process of LLMs and thus have to explore unknown features of LLMs for detection. Instead, watermarks can change the model with a small but pre-defined rule which accelerates the detection process effectively.",
+        "bbox": [
+            507,
+            508,
+            884,
+            831
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Watermark. The distinction between watermark and other methods is that watermark can proactively change the generation to insert a concealed watermark into the generated text. This gives clear difference between watermarked and non",
+        "bbox": [
+            507,
+            841,
+            885,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "614",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/fa529ed5c73dcbd8cd94fb9e4b2e9ab8a405affdc48e4ec8aaa035196bf24ffa.jpg",
+        "image_caption": [
+            "Figure 1: The watermarking process of SemaMark"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            82,
+            884,
+            259
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "watermarked texts. Watermark shifts the text using a small but pre-defined rule to make the detection much more effective. The partition of the vocabulary for each token is a representative watermark method (Kirchenbauer et al., 2023a; Yoo et al., 2023; Kirchenbauer et al., 2023b). In each autoregressive step of generating one token, the method uses the previous tokens' hashes, to select a part of the vocabulary as \"green\" at a ratio of  $\\gamma$ . Subsequently, they elevate the likelihood of the tokens by boosting the logits of the softmax by  $\\delta$ . Through this approach, at each token position, the probability of this matching between the seed and green tokens tends to increase.",
+        "bbox": [
+            112,
+            307,
+            489,
+            533
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "For a sentence with  $L$  tokens, it is viewed as a sample set of size  $L$ . Each token is one sample from the vocabulary. A non-watermarked sentence is expected to have  $\\gamma L$  tokens showing this match, while the watermarked sentence is expected to have more. The watermark detection is approached as a  $z$ -test with null hypothesis that the text is non-watermarke. If the  $z$ -statistic is large, i.e. it is significantly different from the null hypothesis, the null hypothesis can be rejected and the text can be predicted as watermarked:",
+        "bbox": [
+            112,
+            535,
+            489,
+            712
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nz = \\frac {(G - \\gamma L)}{\\sqrt {L \\gamma (1 - \\gamma)}}, \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            226,
+            725,
+            487,
+            762
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $G$  is the number of tokens showing the matching between seed and the green list. Yoo et al. (2023) further expand this watermark of green and red list to more lists for multi-bit encoding.",
+        "bbox": [
+            112,
+            775,
+            487,
+            839
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "(Liu et al., 2023a) propose a semantic invariant method to watermark the generated text of LLM. However, their method employs two additional models, introducing redundant encoding processes in the text encoder, which can be time-consuming.",
+        "bbox": [
+            112,
+            841,
+            489,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Method",
+        "text_level": 1,
+        "bbox": [
+            509,
+            307,
+            613,
+            322
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this section, we introduce the detailed design of SemaMark. We first present how to use the semantic information as the seed for watermark methods that are based on random partition of vocabulary in Section 3.1. Then in Section 3.2 and Section 3.3, we introduce the CL training scheme and the smoothed detection method for further improving the robustness, respectively.",
+        "bbox": [
+            507,
+            336,
+            884,
+            464
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 The framework of SemaMark",
+        "text_level": 1,
+        "bbox": [
+            507,
+            479,
+            791,
+            493
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "As aforementioned, the existing watermark methods based on partition of vocabulary are susceptible to paraphrase. Paraphrase can easily change the previous tokens and disrupt the matching between tokens and the partition of vocabulary, without significantly affecting the semantic meaning. Thus, SemaMark uses the invariant semantics for watermarking by discretizing the embedding space to accommodate the minor perturbation of semantics and provide a stable mapping between semantics and vocabulary partition for the consequent token.",
+        "bbox": [
+            507,
+            502,
+            882,
+            678
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "However, discretization in a high-dimension space is intricate and non-trivial. Therefore, we first reduce the high-dimensional embedding space onto the 2D NE-Ring and then discretize via NE-Ring. The whole watermarking process is shown in Figure 1. SemaMark first reduces the dimension of the embedding space to obtain the discrete semantic values by two steps, i.e., weighted embedding pooling and discretizing by NE-Ring, and then uses the semantic value to partition the vocabulary. The logits of green list is shifted to increase the probability of matching between semantics and the consequent token for watermarking the LLM,  $f$ . In the following, we introduce more details about the two steps to obtain a stable semantic value.",
+        "bbox": [
+            507,
+            680,
+            884,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "615",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "S1: weighted embedding pooling. To enhance the robustness, we aggregate the semantics of previous  $m$  tokens by the weighted mean pooling function  $P(\\cdot)$  before dimension reduction, instead of using only one preceding token's embedding. In the ablation studies of Section 4.4, we show that the method has the best performance when  $m$  is neither too big nor too small. For the token sequence  $\\{t_{i:i + m - 1}\\}$  starting at position  $i$ , we use their semantics to generate the token in the  $m$  position,  $t_{i + m}$ . We denote their embeddings as  $\\{e_{i:i + m - 1}\\}$ .  $\\{e_{i:i + m - 1}\\}$  can be easily obtained from the LLM,  $f$ , that we want to watermark. Intuitively, in  $\\{t_{i:i + m - 1}\\}$ , the embeddings of tokens far from  $t_{i + m}$  contain semantic information that is more distant from  $t_{i + m}$  than the closer ones. The connection between distant tokens might be more possible to change after paraphrase compared with closer tokens. Thus, in the sequence  $\\{t_{i:i + m - 1}\\}$ , the embeddings of distant tokens might be less robust. To increase the robustness for the green list of the current token position  $t_{i + m}$  after paraphrase, the pooling embeddings should rely more on the closer tokens, therefore, we use a linear weight function to assign lower weights to tokens far from  $t_{i + m}$  and higher weights to those in closer proximity:",
+        "bbox": [
+            112,
+            84,
+            489,
+            519
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nP (\\left\\{\\boldsymbol {e} _ {i + 1: i + m} \\right\\}) = \\sum_ {j = 1} ^ {K} \\frac {j + \\frac {K}{2}}{w _ {\\text {s u m}}} \\boldsymbol {e} _ {i + j},\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            169,
+            531,
+            431,
+            576
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $w_{\\mathrm{sum}} = K^2 + K / 2$  is the sum of all weights. We denote the weighted output  $P(\\{e_{i:i + m - 1}\\}) \\in \\mathbb{R}^d$  as  $e_{P_{i,m}}$  for short. By pooling, more semantics are used for a seed, which enhances the robustness under paraphrase.",
+        "bbox": [
+            112,
+            589,
+            487,
+            671
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "S2: discretizing by NE-Ring. After aggregating the embeddings by weighted pooling, SemaMark uses MLP  $g_{\\theta}$  to transform  $e_{P_{i,m}}$  to a normalized vector in 2D embedding space. The normalized vectors locate on a unit circle in the 2D space, which is named as Normalized Embedding Ring (NE-Ring). The discretization function,  $D(\\cdot)$ , discretizes NE-Ring by equally segmenting into different sections. It takes the polar angle  $\\phi$  of  $g_{\\theta}(e_{P_{i,m}})$  as input and outputs the discretized semantic values  $a \\in [K]$ , where  $[K] := \\{1, 2, \\dots, K\\}$ .  $D(\\cdot)$  is defined as",
+        "bbox": [
+            112,
+            681,
+            489,
+            873
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nD (\\phi) = \\left\\lfloor \\phi \\frac {K}{2 \\pi} \\right\\rfloor\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            236,
+            883,
+            361,
+            919
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "It first maps the input from  $[0,2\\pi)$  to  $[0,K)$ , and then discretizes all the values in  $[i,i + 1)$  to  $i$ , for  $\\forall i\\in [K - 1]$ . Even though there could be subtle changes in semantics by paraphrase, the paraphrased  $\\tilde{a}$  will likely locate in the discrete section  $[i,i + 1)$ . Some tokens may still have  $a\\neq \\tilde{a}$  if the normalized vector is close to the boundary of  $[i,i + 1)$ . Therefore, in Section 3.3, we introduce an offset detection to strengthen the tolerance for this mismatch and correct some unstable cases.",
+        "bbox": [
+            507,
+            83,
+            884,
+            243
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "With the two steps, we can get a stable discrete semantic value as the seed for the partition function to partition the vocabulary for the consequent token. Following Kirchenbauer et al. (2023a), the vocabulary is partitioned into green and red lists. We increase the logits of the tokens in the green list by  $\\delta$  and recalculate the probability distribution based on the shifted logits. For each token to generate, we increase the possibility of the green list based on its previous  $m$  tokens' semantics. Thus, all the generated tokens will be likely to have this matching between the semantics and the consequent green token. By detecting the matching, we can discriminate whether a text is watermarked or not and then detect the LLM-generated contents effectively. Besides, SemaMark proposes two strategies to reduce the risk of being cracked by Contrastive Learning and further increase the robustness by the offset detection in the following sections.",
+        "bbox": [
+            507,
+            244,
+            884,
+            551
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2 Training  $g_{\\theta}$  by Contrastive Learning",
+        "text_level": 1,
+        "bbox": [
+            507,
+            562,
+            842,
+            579
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The MLP is expected to produce a uniform distribution of  $g_{\\theta}(e_{P_{i,m}})$  on NE-Ring. If different semantics unevenly distributed on NE-Ring, the resulting discrete semantic values will be overly monotonous and the green list is more changeless. Consequently, the green list might be revealed by counting the token frequency, which compromises the concealment of watermark and leads to the risk of being cracked. Ideally, SemaMark should generate a wider variety of semantic values for different sentences, while each semantic value is robust and stable if its corresponding sentence is paraphrased. To achieve this goal, we propose to use Contrastive Learning to train MLP since Contrastive Learning has the property of uniformity that the data will be evenly distributed in the whole feature space (Wang and Isola, 2020). The uniform distribution can help the normalized vectors cover all the semantic values. As a result, NE-Ring can generate a wider variety of semantic values to prevent the watermark from being cracked.",
+        "bbox": [
+            507,
+            583,
+            884,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "616",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In Contrastive Learning, we first input the sentences into the model  $f$  to get a batch of sequences of  $m$  tokens and their pooling embeddings  $e_{P_{i,m}}$ , denoted as  $\\{e_j\\}$ , where  $j \\in [B]$  and  $B$  is the batch size. To compose a contrastive loss, we construct the positive and negative pairs by a soft augmentation:",
+        "bbox": [
+            112,
+            84,
+            489,
+            197
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\boldsymbol {e} _ {j + B} = \\boldsymbol {e} _ {j} ^ {+} = \\boldsymbol {e} _ {j} + \\epsilon ,\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            215,
+            209,
+            384,
+            229
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\epsilon \\sim \\mathcal{N}(0, \\sigma^2)$  is a Gaussian noise. The soft augmentation can simplify the construction of positive samples. With this soft augmentation, we can assign the samples sharing similar embeddings from the same sequence as positive pairs and samples from different sequences as negative pairs. This is consistent with our intuition that the paraphrased semantic embeddings will not change significantly and can remain robust. Then the contrastive loss is",
+        "bbox": [
+            112,
+            240,
+            489,
+            401
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nL _ {j} = - \\log \\frac {\\exp \\left(\\sin \\left(g _ {\\theta} (\\boldsymbol {e} _ {j}) , g _ {\\theta} (\\boldsymbol {e} _ {j} ^ {+})\\right) / \\tau\\right)}{\\sum_ {k \\neq j , k \\in [ 2 B ]} \\exp \\left(\\sin \\left(g _ {\\theta} (\\boldsymbol {e} _ {j}) , g _ {\\theta} (\\boldsymbol {e} _ {k})\\right) / \\tau\\right)},\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            119,
+            410,
+            480,
+            444
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathrm{sim}(\\cdot)$  is cosine similarity and  $\\tau$  is the temperature. By Contrastive Learning, the output of reduced semantic embeddings can be evenly distributed in all of the space on NE-Ring, and cover all the discrete sections to improve the robustness of SemaMark.",
+        "bbox": [
+            112,
+            454,
+            489,
+            549
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3  $Q$ -offset detection",
+        "text_level": 1,
+        "bbox": [
+            114,
+            561,
+            302,
+            577
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/ab1779953e2c78717d5fb664ef364ddc96d3eaced36bd1c6c77bc584725ffac3.jpg",
+        "image_caption": [
+            "Figure 2:  $Q$ -offset detection vs. existing detection"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            129,
+            594,
+            473,
+            800
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Existing detection methods check the matching between partition seed and the consequent tokens in a one-to-one manner as shown in Figure 2(a). The detection method first recalculates the seed for each token position and gets the partition of the green",
+        "bbox": [
+            112,
+            841,
+            489,
+            921
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "list, and then checks whether the consequent token is in the partitioned green list token by token. In SemaMark, this strategy can be effective when the text is not paraphrased. However, after paraphrase, this detection could be suboptimal because the semantic values of some sequences which are close to the boundaries of the discrete section  $[i,i + 1)$  might change as shown in Figure 2(b). This is because the window of  $m$  tokens will slide token by token during the auto-regressive generation process, and the semantic change will also accumulate when the window is sliding. The semantic values closed to the boundary usually happen when the change accumulates to some extent. This change of boundary semantic values will lead to some mismatch and reduce the accuracy like  $\\tilde{t}_5$  in Figure 2(b).",
+        "bbox": [
+            507,
+            84,
+            884,
+            342
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To mitigate the influence of this error, we propose  $Q$ -offset detection. As shown in Figure 2(c), we offset the discrete seed by  $q$  tokens to detect the matching between semantics and the consequent tokens, where  $q \\in \\{-Q, -(Q - 1), \\dots, 0, 1, \\dots, Q\\}$  and the sign of  $q$  indicates the direction of the offset. We choose the maximal  $z$ -statistic in different  $q$  as the  $Q$ -offset score. However,  $Q$ -offset detection will also increase the  $Q$ -offset score of non-watermark text, which indicates that the detected green word fraction  $\\gamma$  of non-watermark text is higher. The  $\\gamma$  in Eq. (1) is possibly inaccurate. Thus during generation, we set  $\\gamma$  to a fixed value, while in detection process, we treat  $\\gamma$  as a hyperparameter and use a validation set to determine its value in practice. In Section 4.4, we discuss the ablation studies of  $Q$ -offset and  $\\gamma$  and show that  $Q$ -offset can impressively improve the detection performance with robustness.",
+        "bbox": [
+            507,
+            348,
+            884,
+            653
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4 Experiment",
+        "text_level": 1,
+        "bbox": [
+            507,
+            687,
+            648,
+            703
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In this section, we conduct experiments to demonstrate the robustness of SemaMark. In Section 4.2, we demonstrate that its robustness is better than the baseline methods. In Section 4.3, we show that our watermark has almost no influence on the quality of generated texts. In Section 4.4, we use ablation studies to demonstrate the effectiveness of partition function and  $Q$ -offset detection, and show the sensitivity of the partition function. In Section 4.5 we visualize the distribution of NE-Ring and provide analysis on the feature distribution of Contractive Learning.",
+        "bbox": [
+            507,
+            728,
+            884,
+            921
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "617",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/77a7634cfb362a3584e45843c523d649317e2c6811c31d5a993802f08138b69f.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\"></td><td rowspan=\"2\">Paraphrase</td><td colspan=\"4\">ROC-AUC</td><td colspan=\"4\">F1 with best threshold</td></tr><tr><td>LeftHash</td><td>SelfHash</td><td>EXP-Edit</td><td>ours</td><td>LeftHash</td><td>SelfHash</td><td>EXP-Edit</td><td>ours</td></tr><tr><td rowspan=\"4\">OPT-2.7B</td><td>No paraphrase</td><td>0.9913</td><td>0.9886</td><td>0.9799</td><td>0.9948</td><td>0.9921</td><td>0.9861</td><td>0.9708</td><td>0.9905</td></tr><tr><td>Translation</td><td>0.9091</td><td>0.8147</td><td>0.8749</td><td>0.9692</td><td>0.8456</td><td>0.7622</td><td>0.8157</td><td>0.9330</td></tr><tr><td>Dipper</td><td>0.9878</td><td>0.9728</td><td>0.9736</td><td>0.9911</td><td>0.9727</td><td>0.9400</td><td>0.9620</td><td>0.9701</td></tr><tr><td>GPT-3.5</td><td>0.9028</td><td>0.7908</td><td>0.9392</td><td>0.9406</td><td>0.8358</td><td>0.7378</td><td>0.8852</td><td>0.8902</td></tr><tr><td rowspan=\"4\">OPT-6.7B</td><td>No paraphrase</td><td>0.9918</td><td>0.9930</td><td>0.9784</td><td>0.9949</td><td>0.9911</td><td>0.9863</td><td>0.9705</td><td>0.9858</td></tr><tr><td>Translation</td><td>0.8807</td><td>0.8098</td><td>0.8625</td><td>0.9308</td><td>0.8129</td><td>0.7468</td><td>0.8013</td><td>0.8882</td></tr><tr><td>Dipper</td><td>0.9904</td><td>0.9747</td><td>0.9728</td><td>0.9871</td><td>0.9786</td><td>0.9432</td><td>0.9620</td><td>0.9821</td></tr><tr><td>GPT-3.5</td><td>0.8990</td><td>0.7909</td><td>0.8996</td><td>0.9377</td><td>0.8300</td><td>0.7367</td><td>0.8354</td><td>0.8766</td></tr></table>",
+        "bbox": [
+            131,
+            80,
+            867,
+            218
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Watermark detection results under three paraphrases. (The best performance under paraphrase is bolded.)",
+        "bbox": [
+            112,
+            227,
+            880,
+            242
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1 Experiment setups",
+        "text_level": 1,
+        "bbox": [
+            112,
+            267,
+            307,
+            282
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Backbone models and datasets. We test our method by watermarking two models, OPT-2.7B and OPT-6.7B (Zhang et al., 2022) which are referred to as the backbone models in following sections. For dataset, we use the news-like subset of C4 (Raffel et al., 2020), which covers a variety of topics. From the news-like subset of C4, we extract a training set, a validation set and a test set. For each sample, we use the first half of text as prompt to generate watermark sentences. More details can be found in Appendix A.",
+        "bbox": [
+            112,
+            290,
+            487,
+            464
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Baseline methods. We compare our method with three baselines LeftHash, SelfHash (Kirchenbauer et al., 2023b) and EXP-Edit (Kuditipudi et al., 2023). LeftHash and SelfHash are two methods based on the partition of vocabulary using the hashes of tokens. EXP-Edit uses a private sequence to encode the watermark by changing the probability distribution of the sequence of tokens. More details on the implementation can be found in Appendix A.",
+        "bbox": [
+            112,
+            468,
+            487,
+            627
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Paraphrase setups. We use three representative methods to paraphrase the watermarked text, round-trip translation (Tiedemann and Thottingal, 2020), Dipper (Krishna et al., 2023) and GPT-3.5. For round-trip translation, we first translate from English to another language and then transform back to English, such that some words and expressions will be changed because the translation is not an one-to-one mapping. For Dipper, we follow the parameter setting in Kirchenbauer et al. (2023b). For GPT-3.5, we use the prompt in Kirchenbauer et al. (2023b) to query GPT-3.5 for paraphrase. The examples of the three paraphrases can be found at Appendix B.",
+        "bbox": [
+            112,
+            631,
+            487,
+            854
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Evaluation metrics and hyper-parameters. We use F1 score with best threshold and ROC-AUC to measure the performance of the watermark detection. All the metrics are calculated based on",
+        "bbox": [
+            112,
+            857,
+            489,
+            920
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "at least 500 watermarked samples and 500 non-watermark samples. The length of watermarked samples before paraphrase and non-watermark samples is  $200 \\pm 25$ . In generation, we set  $\\gamma = 1/4$  for LeftHash, SelfHash and SemaMark. In detection, we set  $\\gamma = 1/3$  and  $\\delta = 2$  based on the validation set in Section 4.4(b). In SemaMark, we set  $m = 15$ ,  $Q = 15$ ,  $K = 5$  for OPT-2.7B and  $K = 4$  for OPT-6.7B.",
+        "bbox": [
+            507,
+            267,
+            884,
+            411
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2 Main Results",
+        "text_level": 1,
+        "bbox": [
+            507,
+            428,
+            660,
+            442
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In this subsection, we demonstrate the robustness of the proposed SemaMark under paraphrase by comparing it with three baseline methods on two backbone models. We first generate watermarked texts and use three paraphrase methods to remove the watermarks. The detection performance of both texts with and without paraphrase is reported in Table 1. As we can see, before paraphrase, all the watermarked methods have good detection performance. After paraphrase, SemaMark has the best detection performance most of the time across all the backbone models and all the paraphrase methods, which suggests that our method is more robust against paraphrase.",
+        "bbox": [
+            505,
+            451,
+            882,
+            678
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In detail, by round-trip translation, the paraphrase reduces the detection ability of baseline methods effectively, while the watermark of SemaMark is robust. Under round-trip translation, the best ROC-AUC of baselines is 0.9091 on OPT-2.7B and 0.8807 on OPT-6.7B, respectively. But ROC-AUC of SemaMark is 0.9692 and 0.9308, which is at least 0.05 higher than all the baseline methods. Similarly, under paraphrase of GPT-3.5, SemaMark is better than all the baselines. The best baseline performance under GPT-3.5 is 0.9392 in ROC-AUC on OPT-2.7B and 0.8990 in ROC-AUC on opt-6.7B, but SemaMark has higher AUC-ROC of 0.9406 and 0.9377. For Dipper, we note that all methods are robust to Dipper since it does not sig",
+        "bbox": [
+            507,
+            680,
+            884,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "618",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "nificantly reduce the detection performance. However, SemaMark is still one of the most robust. On OPT-2.7B, it performs best in ROC-AUC, while on OPT-6.7B, it has the best F1 score. From Table 1, the results show an obvious improvement of SemaMark in robustness. This implies that using semantics as the seed for the partition function is effective under paraphrase.",
+        "bbox": [
+            112,
+            84,
+            489,
+            212
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.3 Text Quality",
+        "text_level": 1,
+        "bbox": [
+            112,
+            227,
+            258,
+            243
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/30249de8e3a17d6dd17b708f38b4bd694f4ebc5d9e4d1b19b32f8da2539e16f5.jpg",
+        "image_caption": [
+            "(a) OPT-2.7B"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            260,
+            299,
+            391
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/7484c078ce9cb79bd8c654e25b90293de7e490bee8bb54df8866bd5b8fbeb769.jpg",
+        "image_caption": [
+            "(b) OPT-6.7B",
+            "Figure 3: Text quality (perplexity)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            300,
+            261,
+            485,
+            392
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Watermark should not compromise the generation quality of LLMs. In this subsection, we compare the text quality by calculating perplexity and demonstrate that our watermark has almost no influence on the generated quality. Perplexity measures the likelihood that a sentence is generated by one model. Lower perplexity means the watermarked text is more predictable. In other words, it is more consistent with the reasoning of the given model. In Figure 3, we use OPT-6.7B with no watermark to get perplexity for all the watermarked methods. All the results in Figure 3 are calculated without paraphrase, because the generation quality of text is not related to paraphrase. From Figure 3a on OPT-2.7B, we can see that our watermark, LeftHash and SelfHash have almost no influence on the generation quality. They has perplexity at around 6 which is similar as the generated text without watermark. Instead, EXP-Edit has much higher perplexity, which means that EXP-Edit changes the generated text in an aggressive way and much reduces the generation quality after watermarking. This is probably because EXP-Edit adjusts the logits on the whole vocabulary. From Figure 3b, we can draw similar conclusions for OPT-6.7B. EXP-Exit also increases the perplexity by around 10, while the average perplexity of LeftHash, SelfHash and ours is around 1 higher than the non-watermarked generated text. In summary, our SemaMark can",
+        "bbox": [
+            112,
+            454,
+            489,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/f1fedbbb0715eac672aea6cc7255b03e27a0366f52cf5acbf89e20aac0aa7454.jpg",
+        "image_caption": [
+            "Figure 4: ROC-AUC and  $m$"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            559,
+            80,
+            808,
+            211
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/13a767fdb4adaf1727203221315351f0b4c8ead709333794399687afdb5a684b.jpg",
+        "image_caption": [
+            "(a) ROC-AUC and offset  $Q$"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            510,
+            244,
+            684,
+            336
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/1af1c8e4ea05979cb9a711b77fbd619ea31a156df58904c4d9a5119557fb0e15.jpg",
+        "image_caption": [
+            "(b) ROC-AUC and  $\\gamma$",
+            "Figure 5: Text quality (perplexity)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            695,
+            247,
+            863,
+            337
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "keep the quality and robustness simultaneously.",
+        "bbox": [
+            507,
+            406,
+            862,
+            422
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.4 Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            507,
+            432,
+            673,
+            448
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In this subsection, we study the influence of the length of the sequence we use for generating one semantic value and the sensitivity of the partition function.",
+        "bbox": [
+            507,
+            453,
+            880,
+            516
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "a) Length of previous sequence tokens,  $m$ . In the first step of SemaMark, i.e., weighted embedding pooling, we use the semantic of the previous  $m$  tokens to get the more stable embedding. But if the length of the sequence is too long, it will also hurt the robustness. In Figure 4, we test watermark on OPT-2.7B with different  $m$  and draw the ROC-AUC. The results show that before  $m = 15$ , ROC-AUC is in the trend of increase as the  $m$  changes. But when  $m > 15$ , ROC-AUC becomes fluctuating. It is possibly because that the distant tokens will include more change after paraphrase as we mentioned in Section 3.1. Another possible reason is that in the beginning of generation for the first  $m$  tokens, the number of previous tokens is smaller than  $m$  and NE-Ring can only use the embeddings of limited tokens for prediction, which may be unstable. Thus, too long or too short sequence will hurt the robustness of SemaMark against paraphrase. In our experiments, we choose  $m = 15$  for all the settings.",
+        "bbox": [
+            507,
+            527,
+            884,
+            865
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "b)  $Q$ -offset detection In this subsection, we show that the effectiveness of the proposed  $Q$ -offset detection. In Figure 5a, we demonstrate the change",
+        "bbox": [
+            507,
+            873,
+            882,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "619",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "of ROC-AUC of SemaMark with different  $Q$  in offset detection under three different paraphrases.  $Q$ -offset detection searches the highest  $z$ -statistics from  $-Q$  to  $Q$  as the  $Q$ -offset score. From Figure 5a, we can see that when  $Q$  increases, ROC-AUC first increases and decreases after  $Q$  is around 15. When  $Q < 15$ , the offset can help correct the errors of semantic values close to the boundary. Compared with detection without offset, i.e.  $Q = 0$ , ROC-AUC of SemaMark is much better, which means that the offset can help to solve the errors of semantic values around the boundaries that are more vulnerable to paraphrases. When  $Q > 15$ , the correction of this error is limited, because the offset will also increase the  $Q$ -offset score of negative samples as it also searches the highest  $z$ -statistics of negative samples. On the other hand, the computation cost will also increase if  $Q$  is too large because it has to search more possible  $q$ . In practice, we set  $Q = 15$  in all the experiments, which can effectively reduce the influence of the errors of semantic values at the boundaries.",
+        "bbox": [
+            112,
+            84,
+            489,
+            437
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Since the  $Q$ -offset detection searches the highest green word fraction, the fraction of green list word of non-watermarked text will be higher than the  $\\gamma$  that we used to randomly select the green list. Thus, it is not accurate to use the original  $\\gamma$  for  $z$ -statistics. We treat  $\\gamma$  as a hyper-parameter and use a validation set to select its value. As shown in Figure 5b, the detection performance of SemaMark under paraphrases of Dipper and GPT-3.5 will reach the highest when  $\\gamma$  is around  $1/3$ , while it will continue to increase under paraphrase. In practice, we set  $\\gamma = 1/3$  for  $Q$ -offset detection.",
+        "bbox": [
+            112,
+            439,
+            489,
+            633
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "c) Sensitivity of partition function. As we mentioned, the partition function is sensitive to any change of the input because it only uses the input as the seed of the random generator. To validate its sensitivity to continuous embeddings, we adopt the embedding vector as the input to show that, with tiny change of the embeddings, the partition of vocabulary can be very different. We propose a hash method based on md5sum (Deepakumara et al., 2001) to adopt the partition function by transforming the continuous embeddings to an integral seed. We use 1000 sequences to test the sensitivity. For each sequence embedding, we first get a green list from the partition function. Then we change one dimension of the embedding by only 1e-5 to get a new partition result. The overlapping of the green list before and after changing is  $24.99\\%$  on the av",
+        "bbox": [
+            112,
+            646,
+            490,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/3c52149e96e24700b9e9e8ded17e578de606b1d5f764f50fa106f2e01286e666.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\"></td><td colspan=\"3\">ROC-AUC</td><td colspan=\"3\">F1 with best threshold</td></tr><tr><td>LeftHash</td><td>SelfHash</td><td>ours</td><td>LeftHash</td><td>SelfHash</td><td>ours</td></tr><tr><td>LLaMA-7B</td><td>0.819</td><td>0.838</td><td>0.846</td><td>0.748</td><td>0.774</td><td>0.781</td></tr><tr><td>LLaMA2-7B</td><td>0.811</td><td>0.841</td><td>0.872</td><td>0.749</td><td>0.773</td><td>0.810</td></tr></table>",
+        "bbox": [
+            512,
+            80,
+            880,
+            124
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Watermark detection results under different model size.)",
+        "bbox": [
+            507,
+            134,
+            882,
+            162
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "erage of 1000 sequences. It is consistent with  $\\gamma$  we use to watermark, because the random partition with the changed embedding is independent from the original one. It means the partition function is sensitive to any small change in its input. Instead, after we use NE-Ring to discretize the embeddings, the overlapping of green list after changing embeddings by 1e-5 is  $100\\%$ , which means the discretization can effectively handle this change. In practice, SemaMark can provide the tolerance that is much larger than 1e-5, which makes the watermark more robust under paraphrase. With the improvement of  $Q$ -offset, the detection of SemaMark is more robust and effective.",
+        "bbox": [
+            507,
+            190,
+            884,
+            414
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "$d$  ) Model size. To show the robustness of our method on different model sizes, in this section, we also test the watermark under round-trip translation paraphrase on LLaMA-7B and LLaMA2-7B, which have larger size and different architectures. As indicated in Table 3, our approach consistently exhibits the highest robustness against paraphrasing. Specifically, in the LLaMA2-7B model, SemaMark significantly outperforms the baseline models, achieving an increase of 0.06 and 0.03 in ROC-AUC. Similarly, in the LLaMA-7B model, our method shows superior performance with an increase of 0.027 and 0.009 in ROC-AUC.",
+        "bbox": [
+            507,
+            426,
+            885,
+            634
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.5 Distribution on NE-Ring based on CL",
+        "text_level": 1,
+        "bbox": [
+            507,
+            649,
+            853,
+            665
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/ab40d2ca72fb44afadf6354b472756c2c23770329f5cd7f68c9e34bf85f09458.jpg",
+        "image_caption": [
+            "(a) NE-Ring",
+            "Figure 6: Visualization of NE-Ring"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            515,
+            682,
+            690,
+            780
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/c7d99a18669bf41283712fe7dd29af4c993660d36aa002527546083eba306c3a.jpg",
+        "image_caption": [
+            "(b) Distribution on  $\\phi$"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            695,
+            683,
+            875,
+            778
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this subsection, we demonstrate that Contrastive Learning can help evenly distribute the semantics on the NE-Ring. The even distribution can help the sequences reach all possible semantic values and provide more diverse semantic values",
+        "bbox": [
+            507,
+            841,
+            884,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "620",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "to prevent the watermark from being cracked by counting token frequency. In Figure 6a, we use Gaussian density estimation (Chen, 2017) to get the distribution of the semantics on the NE-Ring before discretization. We use different colors to show the density. The NE-Ring in Figure 6a shows that, the distribution is uniform. All the density is between 0.052 and 0.054. We further plot the density based on the polar angle  $\\phi$  in Figure 6b where the density has almost no change on all the polar angle from 0 to  $2\\pi$ . This implies that the training based on Contrastive Learning can ensure the semantics will reach all possible discrete values. It can prevent the case where the discrete values will gather in some discrete sections and produce monotonous vocabulary partitions. As a result, it can protect the watermark from being cracked by counting token frequency.",
+        "bbox": [
+            112,
+            84,
+            489,
+            375
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "5 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            112,
+            387,
+            247,
+            401
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we use the semantic information for watermarking to enhance the robustness against paraphrase. The existing watermark methods use the matching between the previous tokens and the partition vocabulary. This matching can be easily broken by paraphrase. However, we construct the mapping between the semantics and the vocabulary. In this way, the semantics will stay stable under paraphrase and the robustness of watermark can be enhanced. To make use of semantics, we propose SemaMark to discretize the embedding space on NE-Ring and propose a training method based on CL. In addition, we use  $Q$ -offset detection to further advance the robustness by increasing the tolerance of the semantic values close to the discrete boundary. In experiments, we demonstrate our method can perform much better compared with baseline methods under paraphrase while having little influence on the generation quality.",
+        "bbox": [
+            112,
+            414,
+            489,
+            720
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "6 Limitations",
+        "text_level": 1,
+        "bbox": [
+            112,
+            734,
+            250,
+            749
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In some cases, the customers may rely on some API-based LLMs and do not have the access to the embeddings and the permission to modify the logits during generation. Although our watermark method can effectively detect the LLM-generated content and increase the detection success rate under paraphrase, it is not applicable for black-box LLMs. The second weakness of our method is that the NE-Ring is dependent on the semantic embedding of LLMs. For each LLM, we need to train a",
+        "bbox": [
+            112,
+            760,
+            489,
+            921
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "specialized EN-Ring, which might be inflexible if we want to produce a general model for NE-Ring or fine-tune the LLMs. Despite the weaknesses, our method is successful in the problem of robustness under paraphrase. In the future work, we will continue to extent our method into black-box LLMs and a universal model that does not require customized training for various specific LLMs.",
+        "bbox": [
+            507,
+            84,
+            884,
+            212
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Potential risk. Our discussion about the robustness might provide motivation for the attackers to find other methods like adaptive attack. Although we provide robustness under paraphrase, if the unauthorized people propose possible attack method focusing on the green-list based watermark from other perspectives, the detection rate for LLM-generated texts are still possible to be reduced.",
+        "bbox": [
+            507,
+            212,
+            885,
+            341
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "7 Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            507,
+            353,
+            712,
+            370
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Jie Ren, Han Xu, Yingqian Cui, and Jiliang Tang are supported by the National Science Foundation (NSF) under grant numbers CNS 2246050, IIS1845081, IIS2212032, IIS2212144, IOS2107215, DUE 2234015, DRL 2025244 and IOS2035472, the Army Research Office (ARO) under grant number W911NF-21-1-0198, the Home Depot, Cisco Systems Inc, Amazon Faculty Award, Johnson&Johnson, JP Morgan Faculty Award and SNAP.",
+        "bbox": [
+            507,
+            379,
+            884,
+            539
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            509,
+            565,
+            608,
+            581
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Steven Bird, Ewan Klein, and Edward Loper. 2009. Natural language processing with Python: analyzing text with the natural language toolkit. \"O'Reilly Media, Inc.\"",
+            "Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR.",
+            "Yen-Chi Chen. 2017. A tutorial on kernel density estimation and recent advances. Biostatistics & Epidemiology, 1(1):161-187.",
+            "Antonia Creswell, Murray Shanahan, and Irina Higgins. 2022. Selection-inference: Exploiting large language models for interpretable logical reasoning. arXiv preprint arXiv:2205.09712.",
+            "Janaka Deepakumara, Howard M Heys, and R Venkatesan. 2001. FPGA implementation of md5 hash algorithm. In Canadian Conference on Electrical and Computer Engineering 2001. Conference Proceedings (Cat. No. 01TH8555), volume 2, pages 919-924. IEEE."
+        ],
+        "bbox": [
+            509,
+            589,
+            885,
+            919
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "621",
+        "bbox": [
+            485,
+            928,
+            512,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Tianyu Gao, Xingcheng Yao, and Danqi Chen. 2021. Simcse: Simple contrastive learning of sentence embeddings. In 2021 Conference on Empirical Methods in Natural Language Processing, EMNLP 2021, pages 6894-6910. Association for Computational Linguistics (ACL).",
+            "John Giorgi, Osvald Nitski, Bo Wang, and Gary Bader. 2021. Declutr: Deep contrastive learning for unsupervised textual representations. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 879-895.",
+            "Biyang Guo, Xin Zhang, Ziyuan Wang, Minqi Jiang, Jinran Nie, Yuxuan Ding, Jianwei Yue, and Yupeng Wu. 2023. How close is chatgpt to human experts? comparison corpus, evaluation, and detection. arXiv preprint arXiv:2301.07597.",
+            "Charles R. Harris, K. Jarrod Millman, Stefan J. van der Walt, Ralf Gommers, Pauli Virtanen, David Cournaepau, Eric Wieser, Julian Taylor, Sebastian Berg, Nathaniel J. Smith, Robert Kern, Matti Picus, Stephan Hoyer, Marten H. van Kerkwijk, Matthew Brett, Allan Haldane, Jaime Fernandez del Rio, Mark Wiebe, Pearu Peterson, Pierre Gerard-Marchant, Kevin Sheppard, Tyler Reddy, Warren Weckesser, Hameer Abbasi, Christoph Gohlke, and Travis E. Oliphant. 2020. Array programming with NumPy. Nature, 585(7825):357-362.",
+            "Jean Kaddour, Joshua Harris, Maximilian Mozes, Herb Bradley, Roberta Raileanu, and Robert McHardy. 2023. Challenges and applications of large language models. arXiv preprint arXiv:2307.10169.",
+            "John Kirchenbauer, Jonas Geiping, Yuxin Wen, Jonathan Katz, Ian Miers, and Tom Goldstein. 2023a. A watermark for large language models. arXiv preprint arXiv:2301.10226.",
+            "John Kirchenbauer, Jonas Geiping, Yuxin Wen, Manli Shu, Khalid Saifullah, Kezhi Kong, Kasun Fernando, Aniruddha Saha, Micah Goldblum, and Tom Goldstein. 2023b. On the reliability of watermarks for large language models. arXiv preprint arXiv:2306.04634.",
+            "Sarah Kreps, R Miles McCain, and Miles Brundage. 2022. All the news that's fit to fabricate: A-generated text as a tool of media misinformation. Journal of experimental political science, 9(1):104-117.",
+            "Kalpesh Krishna, Yixiao Song, Marzena Karpinska, John Wieting, and Mohit Iyyer. 2023. Paraphrasing evades detectors of ai-generated text, but retrieval is an effective defense. arXiv preprint arXiv:2303.13408.",
+            "Rohith Kuditipudi, John Thickstun, Tatsunori Hashimoto, and Percy Liang. 2023. Robust distortion-free watermarks for language models. arXiv preprint arXiv:2307.15593."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            919
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Bohan Li, Hao Zhou, Junxian He, Mingxuan Wang, Yiming Yang, and Lei Li. 2020. On the sentence embeddings from pre-trained language models. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 9119-9130.",
+            "Linyang Li, Pengyu Wang, Ke Ren, Tianxiang Sun, and Xipeng Qiu. 2023. Origin tracing and detecting of llms. arXiv preprint arXiv:2304.14072.",
+            "Aiwei Liu, Leyi Pan, Xuming Hu, Shiao Meng, and Lijie Wen. 2023a. A semantic invariant robust watermark for large language models. arXiv preprint arXiv:2310.06356.",
+            "Aiwei Liu, Leyi Pan, Yijian Lu, Jingjing Li, Xuming Hu, Lijie Wen, Irwin King, and Philip S Yu. 2023b. A survey of text watermarking in the era of large language models. arXiv preprint arXiv:2312.07913.",
+            "Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. 2022. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521.",
+            "Eric Mitchell, Yoonho Lee, Alexander Khazatsky, Christopher D Manning, and Chelsea Finn. 2023. Detectgpt: Zero-shot machine-generated text detection using probability curvature. arXiv preprint arXiv:2301.11305.",
+            "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67.",
+            "Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3982-3992.",
+            "Jie Ren, Han Xu, Pengfei He, Yingqian Cui, Shenglai Zeng, Jiankun Zhang, Hongzhi Wen, Jiayuan Ding, Hui Liu, Yi Chang, et al. 2024. Copyright protection in generative ai: A technical perspective. arXiv preprint arXiv:2402.02333.",
+            "Vinu Sankar Sadasivan, Aounon Kumar, Sriram Balasubramanian, Wenxiao Wang, and Soheil Feizi. 2023. Can ai-generated text be reliably detected? arXiv preprint arXiv:2303.11156.",
+            "Joni Salminen, Chandrashekhar Kandpal, Ahmed Mohamed Kamel, Soon-gyo Jung, and Bernard J Jansen. 2022. Creating and detecting fake reviews of online products. Journal of Retailing and Consumer Services, 64:102771."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            919
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "622",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Ruixiang Tang, Yu-Neng Chuang, and Xia Hu. 2023. The science of detecting llm-generated texts. arXiv preprint arXiv:2303.07205.",
+            "Jörg Tiedemann and Santhosh Thottingal. 2020. OPUSMT — Building open translation services for the World. In Proceedings of the 22nd Annual Conference of the European Association for Machine Translation (EAMT), Lisbon, Portugal.",
+            "Saranya Venkatraman, Adaku Uchendu, and Dongwon Lee. 2023. Gpt-who: An information density-based machine-generated text detector.",
+            "Pengyu Wang, Linyang Li, Ke Ren, Botian Jiang, Dong Zhang, and Xipeng Qiu. 2023. Seqxgpt: Sentence-level ai-generated text detection.",
+            "Tongzhou Wang and Phillip Isola. 2020. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International Conference on Machine Learning, pages 9929-9939. PMLR.",
+            "Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35:24824-24837.",
+            "Junchao Wu, Shu Yang, Runzhe Zhan, Yulin Yuan, Derek F Wong, and Lidia S Chao. 2023. A survey on llm-gernerated text detection: Necessity, methods, and future directions. arXiv preprint arXiv:2310.14724.",
+            "Frank F Xu, Uri Alon, Graham Neubig, and Vincent Josua Hellendoorn. 2022. A systematic evaluation of large language models of code. In Proceedings of the 6th ACM SIGPLAN International Symposium on Machine Programming, pages 1-10.",
+            "Han Xu, Jie Ren, Pengfei He, Shenglai Zeng, Yingqian Cui, Amy Liu, Hui Liu, and Jiliang Tang. 2023. On the generalization of training-based chatgpt detection methods. arXiv preprint arXiv:2310.01307.",
+            "KiYoon Yoo, Wonhyuk Ahn, Jiho Jang, and Nojun Kwak. 2023. Robust multi-bit natural language watermarking through invariant features. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2092-2115.",
+            "Rowan Zellers, Ari Holtzman, Hannah Rashkin, Yonatan Bisk, Ali Farhadi, Franziska Roesner, and Yejin Choi. 2019. Defending against neural fake news. Advances in neural information processing systems, 32.",
+            "Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. 2022. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068."
+        ],
+        "bbox": [
+            115,
+            85,
+            485,
+            919
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "ref_text",
+        "text": "Xuandong Zhao, Prabhanjan Ananth, Lei Li, and Yu-Xiang Wang. 2023. Provable robust watermarking for ai-generated text. arXiv preprint arXiv:2306.17439.",
+        "bbox": [
+            510,
+            85,
+            880,
+            137
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "623",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A More details on experimental settings",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            473,
+            99
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "All the baseline models, backbone models and datasets we use are open source and available for academic purpose. For backbone models, we use the open-sourced model from Huggingface<sup>1</sup>. The implementation is based on Pytorch<sup>2</sup> framework and also depend on packages including NLTK (Bird et al., 2009) and Numpy (Harris et al., 2020). For baseline methods, we use the released official code from the authors. For paraphrase models, we use OPUS-MT translation model and Dipper on Huggingface repository<sup>3</sup>, and API of ChatGPT<sup>4</sup>.",
+        "bbox": [
+            112,
+            109,
+            489,
+            300
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B Examples of paraphrases",
+        "text_level": 1,
+        "bbox": [
+            112,
+            313,
+            369,
+            330
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "In Table 3, we provide the examples of three paraphrases.",
+        "bbox": [
+            112,
+            338,
+            489,
+            370
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{1}$ https://huggingface.co/facebook/opt-2.7b",
+        "bbox": [
+            137,
+            856,
+            394,
+            869
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{2}$ https://pytorch.org/",
+        "bbox": [
+            137,
+            869,
+            263,
+            882
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>3</sup>https://huggingface.co/Helsinki-NLP/opus-mt-en-zh and https://huggingface.co/kalpeshk2011/dipper-paraphraser-xxl",
+        "bbox": [
+            117,
+            882,
+            485,
+            906
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>4</sup>https://chat.openai.com/chat",
+        "bbox": [
+            137,
+            906,
+            315,
+            920
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "624",
+        "bbox": [
+            485,
+            928,
+            515,
+            939
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/c6be4b2387c0ce76b6c949388684e26831312b7efab3e4dfe48db46372fb0579.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Original</td><td>Paraphrased</td></tr><tr><td>Translation</td><td>The course ran from Feb. 16 to Feb. 18 and was designed to address officers&#x27; response to critical incidents and to enhance officer safety. &quot;I am excited by the training and professionalism demonstrated by the officers at the training session. They were receptive and focused on learning new techniques and approaches to policing.&quot; Com-missioner Walwyn added.</td><td>The training, which was held from 16 February to 18 February, was aimed at discussing police responses to major incidents and enhancing police security. &quot; I am excited about the training and professionalism shown by police officers in the course, who are willing to accept and focus on new policing techniques and methods. &quot; Com-missioner Walwyn adds.</td></tr><tr><td>GPT</td><td>She tells him she wants to protect him. She asks what else she can do. Stefan tells her not to trust Chad. She vows she won&#x27;t. Stefan offers to get rid of Chad for her if she wants. She says she wants Chad and his family gone. She wants to have his babies. She tells him to get out of her way. She doesn&#x27;t know if she can love him, but she wants his babies. She wants to be a mother. Stefan tells her she deserves a real man. He&#x27;ll love her and protect her. She says she wants him and his babies. She wants to have his babies. Stefan says she deserves a real man. He&#x27;ll love her and protect her. She says she wants him and his babies. She wants to have his babies. Stefan says she deserves a real man. He&#x27;ll love her and protect her. She says she wants him and his babies. She wants to have his babies. Stefan says she deserves a real man. He&#x27;ll love her and protect her. She says she wants him and his babies.</td><td>According to Reuters, China has already taken steps to solidify its presence in the South China Sea. It has constructed an airstrip on one island and In a tense conversation, a woman expresses her desire to protect a man, Stefan, and seeks his advice on how to do so. Stefan advises her not to trust a person named Chad, to which she firmly pledges her distrust. Stefan offers to eliminate Chad from her life if she desires, and she responds by expressing her wish for Chad and his family to disappear entirely. In a surprising revelation, she declares her desire to have Stefan&#x27;s children, emphasizing her longing to become a mother. Despite uncertain feelings of love towards Stefan, she insists on wanting his babies and urges him to step aside. Stefan, in turn, assures her that she deserves a genuine partner who will love and safeguard her. Yet, undeterred, she asserts her desire for both Stefan and his offspring, reiterat-ing her ambition to have his babies.</td></tr><tr><td>DIPPER</td><td>Kewanee got the ball again with 7.2 seconds to go, but couldn&#x27;t get a shot off. &quot;I think it was something we needed to have a plan for and it was one that worked in the end. We did it well enough,&quot; Keene said. &quot;We were trying to get to the rim. A lot of times when you are in a situation like that and you can&#x27;t score, it&#x27;s better to just foul.&quot; Macomb was 12-for-27 from behind the arc (42 percent). The Grinnell offense is centered around perimeter shooters. The two perimeter players for Macomb are Emerick and Reilly Rieder. The next closest shooter on the team is Rieder with 10 points. Nolan led Kewanee with 15 points. Macomb, now 14-5, now travels to Jacksonville on Thursday to play Jacksonville</td><td>Kewanee had the ball again with 7.2 seconds left but couldn&#x27;t get off a shot. &quot;I think that&#x27;s something we needed to have a plan for and in the end, it worked,&quot; coach Lambert said. &quot;We did it well enough,&quot; Keene said. &quot;We were trying to get to the rim. Often when you are in a situation like that and you can&#x27;t score, it&#x27;s better to foul.&quot; Macomb was 12 for 27 from beyond the arc (42 percent). The Grinnell offense is based on sharpshooting players. Macomb&#x27;s two shooters are Emerick and Rieder. Rieder has ten points. Nolan led Kewanee with 15 points. Macomb, now 14-5, will play at Jacksonville Thursday.</td></tr></table>",
+        "bbox": [
+            137,
+            105,
+            858,
+            868
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Paraphrase examples.",
+        "bbox": [
+            389,
+            876,
+            603,
+            891
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "625",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 12
+    }
+]

2024/A Robust Semantics-based Watermark for Large Language Model against Paraphrasing/b1276659-b3bd-43b9-95dd-a7cb02ea87a5_model.json

@@ -0,0 +1,1942 @@
+[
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.151,
+                0.09,
+                0.848,
+                0.131
+            ],
+            "angle": 0,
+            "content": "A Robust Semantics-based Watermark for Large Language Models against Paraphrasing"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.213,
+                0.146,
+                0.79,
+                0.181
+            ],
+            "angle": 0,
+            "content": "Jie Ren\\(^{1}\\), Han Xu\\(^{1}\\), Yiding Liu\\(^{2}\\), Yingqian Cui\\(^{1}\\), Shuaiqiang Wang\\(^{2}\\), Dawei Yin\\(^{2}\\), Jiliang Tang\\(^{1}\\)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.342,
+                0.181,
+                0.662,
+                0.197
+            ],
+            "angle": 0,
+            "content": "<sup>1</sup>Michigan State University, <sup>2</sup>Baidu Inc."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.273,
+                0.199,
+                0.729,
+                0.214
+            ],
+            "angle": 0,
+            "content": "{renjie3, xuhan1, cuiyingq, tangjili}@msu.edu"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.168,
+                0.215,
+                0.835,
+                0.23
+            ],
+            "angle": 0,
+            "content": "liuyiding.tanh@gmail.com, shqiang.wang@gmail.com, yindawei@acm.org"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.261,
+                0.262,
+                0.341,
+                0.276
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.145,
+                0.284,
+                0.461,
+                0.681
+            ],
+            "angle": 0,
+            "content": "Large language models (LLMs) have show their remarkable ability in various natural language tasks. However, there are concerns that LLMs are possible to be used improperly or even illegally. To prevent the malicious usage of LLMs, detecting LLM-generated text becomes crucial in the deployment of LLM applications. Watermarking is an effective strategy to detect the LLM-generated content by encoding a pre-defined secret watermark to facilitate the detection process. However, the majority of existing watermark methods leverage the simple hashes of precedent tokens to partition vocabulary. Such watermarks can be easily eliminated by paraphrase and, correspondingly, the detection effectiveness will be greatly compromised. Thus, to enhance the robustness against paraphrase, we propose a semantics-based watermark framework, SemaMark. It leverages the semantics as an alternative to simple hashes of tokens since the semantic meaning of the sentences will be likely preserved under paraphrase and the watermark can remain robust. Comprehensive experiments are conducted to demonstrate the effectiveness and robustness of SemaMark under different paraphrases. Our code is available at github.com/renjie3/SemaMark."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.688,
+                0.26,
+                0.703
+            ],
+            "angle": 0,
+            "content": "1 Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.713,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Large language models (LLMs) have shown their great ability in various natural language processing (NLP) tasks like Question Answering (QA) (Lu et al., 2022), reasoning tasks (Wei et al., 2022; Creswell et al., 2022) and code development (Xu et al., 2022). However, tremendous concerns have been raised that LLMs are possible to be used improperly and illegally. For example, indistinguishable fake news are easy to be fabricated (Kreps et al., 2022; Zellers et al., 2019) by language models, which, when disseminated, could instigate widespread panic. Similarly, in the commercial sphere, convincingly generated reviews"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.262,
+                0.885,
+                0.357
+            ],
+            "angle": 0,
+            "content": "can manipulate consumer perceptions, leading to unethical business competition (Salminen et al., 2022). Therefore, detecting LLM-generated text has become crucial in the real-world applications of LLMs (Wu et al., 2023; Sadasivan et al., 2023; Xu et al., 2023)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.359,
+                0.885,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Among diverse methods to detect LLM-generated texts, the watermark strategies have demonstrated outstanding precision (Liu et al., 2023b; Tang et al., 2023; Ren et al., 2024). It is proposed to encode a secret watermark into the generated texts, such that we can tell whether a text is generated by detecting this watermark. One representative strategy (Kirchenbauer et al., 2023a; Yoo et al., 2023) is to encode the watermark based on the \"partition of vocabulary\". In detail, given a language model, these methods devise a mapping from precedent tokens to a particular partition of the vocabulary by a partition function for the consequent token. The partition function leverages the hashes of the input as the seed of a random generator to split the vocabulary to a green list and a red list. During the text generation phase, the consequent token has an increased probability to be sampled from the green list. In this way, the watermark is encoded through the matching between the precedent tokens and the vocabulary partition for the consequent token. The detection is also facilitated by detecting this matching in generated contents. However, recent works (Krishna et al., 2023; Kaddour et al., 2023) reveal that this watermark may be easily eliminated by sentence paraphrasing. Individuals seeking to improperly utilize LLMs without being detected can paraphrase the generated contents, like altering the order and the choices of the words, and only retain the general meaning of the text to achieve their malicious goals like faking news. These paraphrases will change the seed of the partition function, i.e. the token hashes, and as we show in the Section 4.4, the partition function is sensitive to small changes. Consequently,"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.486,
+                0.929,
+                0.516,
+                0.941
+            ],
+            "angle": 0,
+            "content": "613"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.228,
+                0.946,
+                0.77,
+                0.973
+            ],
+            "angle": 0,
+            "content": "Findings of the Association for Computational Linguistics: NAACL 2024, pages 613-625 June 16-21, 2024 ©2024 Association for Computational Linguistics"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.489,
+                0.149
+            ],
+            "angle": 0,
+            "content": "the matching between the precedent tokens and the green list will be disrupted, and the detection effectiveness of the watermark can be dramatically compromised."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.151,
+                0.49,
+                0.47
+            ],
+            "angle": 0,
+            "content": "In this paper, we propose to leverage the semantic meaning of precedent token sequences as the seed for partition function, instead of simple hashes of precedent tokens, since the core semantic meaning is expected to be maintained after paraphrase. To achieve this goal, one key obstacle is how to capture the semantics when applying them for the partition function to watermark the generated texts. It is a common practice to quantify the semantics via embeddings (Reimers and Gurevych, 2019; Gao et al., 2021; Li et al., 2020; Giorgi et al., 2021). Embeddings indeed can represent consistent semantics after paraphrase. Since the embeddings are high-dimensional vectors in the continuous space, they often present some minor changes after paraphrase. Although the main semantics are preserved, these minor changes can lead to a substantial difference in the partition of vocabulary because the random generator in the partition function is sensitive to the change of the seed, as shown in Section 4.4."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.472,
+                0.49,
+                0.921
+            ],
+            "angle": 0,
+            "content": "To overcome the above challenge, i.e., to make the quantified semantics invariant and make the watermark robust under paraphrase, we propose a new watermark method, SemaMark, which discretizes the continuous embedding space. Intuitively, the discretization can coarsen the representation of the embeddings which could tolerate the potential minor changes caused by paraphrase. By proper discretization, the paraphrased semantics could stay in the same discrete section with a high probability and the discretized quantified semantics will likely remain the same even after paraphrase. Therefore, the partition results will not change. However, directly converting the high-dimensional embedding space into discrete is intricate and challenging. For example, discretizing each dimension will lead to a large amount of discrete values which is exponential to the number of dimensions. Thus, the minor changes by paraphrase can still cause the change of discrete values because the number of discrete values are too dense and each discrete value can tolerate only small changes. Therefore, the minor changes of high-dimensional embeddings can have a strong impact on the partition function. To address this problem, SemaMark first uses a MultiLayer Perception (MLP) to condense the continuous high-dimensional embeddings into normalized vectors in 2D space. The vectors are located"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.885,
+                0.472
+            ],
+            "angle": 0,
+            "content": "on a unit circle named Normalized Embedding Ring (NE-Ring). Then the condensed NE-Ring is equally divided into various sections, transforming the continuous space into distinct discrete values, i.e., \"semantic values\". Based on the discretization, SemaMark further introduces two strategies to advance the watermark's concealment and to improve the robustness under paraphrase. First, SemaMark leverages the uniformity (Wang and Isola, 2020) of Contrastive Learning(CL) (Chen et al., 2020) to strength the MLP and mitigate the problem that the semantics are unevenly concentrating on some discrete sections on NE-Ring. The unevenly distribution will cause the final discrete semantic values overly monotonous. It raises the concern that the watermark might be cracked by counting token frequency (Zhao et al., 2023). Second, SemaMark utilizes an offset detection method to further enhance the robustness at the boundary of different discrete sections whose semantic values are possibly vulnerable to paraphrase. Comprehensive experiments are conducted to demonstrate the effectiveness and robustness of SemaMark under different paraphrases."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.484,
+                0.671,
+                0.498
+            ],
+            "angle": 0,
+            "content": "2 Related works"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.51,
+                0.885,
+                0.832
+            ],
+            "angle": 0,
+            "content": "LLM-generated detection. As the development of LLMs, various LLM-generated detection tools have also been proposed. Learning-based methods train a classification model to detect the difference between human-written text and machine-generated text like Guo et al. (2023); Wang et al. (2023); Li et al. (2023). Other works do not rely on the classification model, but try to use the property of the LLM to test whether a given text is generated by LLMs. For example, DetectGPT (Mitchell et al., 2023) assumes that the generated text will have high likelihood. GPT-who (Venkatraman et al., 2023) uses UID-based features to model the unique statistical signature of each LLM and human author for accurate authorship attribution. These methods do not interact the generation process of LLMs and thus have to explore unknown features of LLMs for detection. Instead, watermarks can change the model with a small but pre-defined rule which accelerates the detection process effectively."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.842,
+                0.887,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Watermark. The distinction between watermark and other methods is that watermark can proactively change the generation to insert a concealed watermark into the generated text. This gives clear difference between watermarked and non"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.486,
+                0.929,
+                0.517,
+                0.941
+            ],
+            "angle": 0,
+            "content": "614"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.118,
+                0.083,
+                0.885,
+                0.26
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.325,
+                0.269,
+                0.673,
+                0.285
+            ],
+            "angle": 0,
+            "content": "Figure 1: The watermarking process of SemaMark"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.309,
+                0.49,
+                0.534
+            ],
+            "angle": 0,
+            "content": "watermarked texts. Watermark shifts the text using a small but pre-defined rule to make the detection much more effective. The partition of the vocabulary for each token is a representative watermark method (Kirchenbauer et al., 2023a; Yoo et al., 2023; Kirchenbauer et al., 2023b). In each autoregressive step of generating one token, the method uses the previous tokens' hashes, to select a part of the vocabulary as \"green\" at a ratio of \\(\\gamma\\). Subsequently, they elevate the likelihood of the tokens by boosting the logits of the softmax by \\(\\delta\\). Through this approach, at each token position, the probability of this matching between the seed and green tokens tends to increase."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.536,
+                0.49,
+                0.713
+            ],
+            "angle": 0,
+            "content": "For a sentence with \\( L \\) tokens, it is viewed as a sample set of size \\( L \\). Each token is one sample from the vocabulary. A non-watermarked sentence is expected to have \\( \\gamma L \\) tokens showing this match, while the watermarked sentence is expected to have more. The watermark detection is approached as a \\( z \\)-test with null hypothesis that the text is non-watermarke. If the \\( z \\)-statistic is large, i.e. it is significantly different from the null hypothesis, the null hypothesis can be rejected and the text can be predicted as watermarked:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.227,
+                0.726,
+                0.488,
+                0.763
+            ],
+            "angle": 0,
+            "content": "\\[\nz = \\frac {(G - \\gamma L)}{\\sqrt {L \\gamma (1 - \\gamma)}}, \\tag {1}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.776,
+                0.489,
+                0.84
+            ],
+            "angle": 0,
+            "content": "where \\(G\\) is the number of tokens showing the matching between seed and the green list. Yoo et al. (2023) further expand this watermark of green and red list to more lists for multi-bit encoding."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.842,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": "(Liu et al., 2023a) propose a semantic invariant method to watermark the generated text of LLM. However, their method employs two additional models, introducing redundant encoding processes in the text encoder, which can be time-consuming."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.308,
+                0.614,
+                0.323
+            ],
+            "angle": 0,
+            "content": "3 Method"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.337,
+                0.885,
+                0.466
+            ],
+            "angle": 0,
+            "content": "In this section, we introduce the detailed design of SemaMark. We first present how to use the semantic information as the seed for watermark methods that are based on random partition of vocabulary in Section 3.1. Then in Section 3.2 and Section 3.3, we introduce the CL training scheme and the smoothed detection method for further improving the robustness, respectively."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.48,
+                0.792,
+                0.494
+            ],
+            "angle": 0,
+            "content": "3.1 The framework of SemaMark"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.503,
+                0.884,
+                0.679
+            ],
+            "angle": 0,
+            "content": "As aforementioned, the existing watermark methods based on partition of vocabulary are susceptible to paraphrase. Paraphrase can easily change the previous tokens and disrupt the matching between tokens and the partition of vocabulary, without significantly affecting the semantic meaning. Thus, SemaMark uses the invariant semantics for watermarking by discretizing the embedding space to accommodate the minor perturbation of semantics and provide a stable mapping between semantics and vocabulary partition for the consequent token."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.681,
+                0.885,
+                0.922
+            ],
+            "angle": 0,
+            "content": "However, discretization in a high-dimension space is intricate and non-trivial. Therefore, we first reduce the high-dimensional embedding space onto the 2D NE-Ring and then discretize via NE-Ring. The whole watermarking process is shown in Figure 1. SemaMark first reduces the dimension of the embedding space to obtain the discrete semantic values by two steps, i.e., weighted embedding pooling and discretizing by NE-Ring, and then uses the semantic value to partition the vocabulary. The logits of green list is shifted to increase the probability of matching between semantics and the consequent token for watermarking the LLM, \\( f \\). In the following, we introduce more details about the two steps to obtain a stable semantic value."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.486,
+                0.929,
+                0.516,
+                0.941
+            ],
+            "angle": 0,
+            "content": "615"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.49,
+                0.52
+            ],
+            "angle": 0,
+            "content": "S1: weighted embedding pooling. To enhance the robustness, we aggregate the semantics of previous \\( m \\) tokens by the weighted mean pooling function \\( P(\\cdot) \\) before dimension reduction, instead of using only one preceding token's embedding. In the ablation studies of Section 4.4, we show that the method has the best performance when \\( m \\) is neither too big nor too small. For the token sequence \\( \\{t_{i:i + m - 1}\\} \\) starting at position \\( i \\), we use their semantics to generate the token in the \\( m \\) position, \\( t_{i + m} \\). We denote their embeddings as \\( \\{e_{i:i + m - 1}\\} \\). \\( \\{e_{i:i + m - 1}\\} \\) can be easily obtained from the LLM, \\( f \\), that we want to watermark. Intuitively, in \\( \\{t_{i:i + m - 1}\\} \\), the embeddings of tokens far from \\( t_{i + m} \\) contain semantic information that is more distant from \\( t_{i + m} \\) than the closer ones. The connection between distant tokens might be more possible to change after paraphrase compared with closer tokens. Thus, in the sequence \\( \\{t_{i:i + m - 1}\\} \\), the embeddings of distant tokens might be less robust. To increase the robustness for the green list of the current token position \\( t_{i + m} \\) after paraphrase, the pooling embeddings should rely more on the closer tokens, therefore, we use a linear weight function to assign lower weights to tokens far from \\( t_{i + m} \\) and higher weights to those in closer proximity:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.17,
+                0.532,
+                0.433,
+                0.577
+            ],
+            "angle": 0,
+            "content": "\\[\nP (\\left\\{\\boldsymbol {e} _ {i + 1: i + m} \\right\\}) = \\sum_ {j = 1} ^ {K} \\frac {j + \\frac {K}{2}}{w _ {\\text {s u m}}} \\boldsymbol {e} _ {i + j},\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.59,
+                0.489,
+                0.673
+            ],
+            "angle": 0,
+            "content": "where \\( w_{\\mathrm{sum}} = K^2 + K / 2 \\) is the sum of all weights. We denote the weighted output \\( P(\\{e_{i:i + m - 1}\\}) \\in \\mathbb{R}^d \\) as \\( e_{P_{i,m}} \\) for short. By pooling, more semantics are used for a seed, which enhances the robustness under paraphrase."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.682,
+                0.49,
+                0.875
+            ],
+            "angle": 0,
+            "content": "S2: discretizing by NE-Ring. After aggregating the embeddings by weighted pooling, SemaMark uses MLP \\( g_{\\theta} \\) to transform \\( e_{P_{i,m}} \\) to a normalized vector in 2D embedding space. The normalized vectors locate on a unit circle in the 2D space, which is named as Normalized Embedding Ring (NE-Ring). The discretization function, \\( D(\\cdot) \\), discretizes NE-Ring by equally segmenting into different sections. It takes the polar angle \\( \\phi \\) of \\( g_{\\theta}(e_{P_{i,m}}) \\) as input and outputs the discretized semantic values \\( a \\in [K] \\), where \\( [K] := \\{1, 2, \\dots, K\\} \\). \\( D(\\cdot) \\) is defined as"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.237,
+                0.884,
+                0.363,
+                0.92
+            ],
+            "angle": 0,
+            "content": "\\[\nD (\\phi) = \\left\\lfloor \\phi \\frac {K}{2 \\pi} \\right\\rfloor\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.084,
+                0.885,
+                0.244
+            ],
+            "angle": 0,
+            "content": "It first maps the input from \\([0,2\\pi)\\) to \\([0,K)\\), and then discretizes all the values in \\([i,i + 1)\\) to \\(i\\), for \\(\\forall i\\in [K - 1]\\). Even though there could be subtle changes in semantics by paraphrase, the paraphrased \\(\\tilde{a}\\) will likely locate in the discrete section \\([i,i + 1)\\). Some tokens may still have \\(a\\neq \\tilde{a}\\) if the normalized vector is close to the boundary of \\([i,i + 1)\\). Therefore, in Section 3.3, we introduce an offset detection to strengthen the tolerance for this mismatch and correct some unstable cases."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.246,
+                0.885,
+                0.552
+            ],
+            "angle": 0,
+            "content": "With the two steps, we can get a stable discrete semantic value as the seed for the partition function to partition the vocabulary for the consequent token. Following Kirchenbauer et al. (2023a), the vocabulary is partitioned into green and red lists. We increase the logits of the tokens in the green list by \\(\\delta\\) and recalculate the probability distribution based on the shifted logits. For each token to generate, we increase the possibility of the green list based on its previous \\(m\\) tokens' semantics. Thus, all the generated tokens will be likely to have this matching between the semantics and the consequent green token. By detecting the matching, we can discriminate whether a text is watermarked or not and then detect the LLM-generated contents effectively. Besides, SemaMark proposes two strategies to reduce the risk of being cracked by Contrastive Learning and further increase the robustness by the offset detection in the following sections."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.563,
+                0.843,
+                0.58
+            ],
+            "angle": 0,
+            "content": "3.2 Training \\(g_{\\theta}\\) by Contrastive Learning"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.584,
+                0.885,
+                0.922
+            ],
+            "angle": 0,
+            "content": "The MLP is expected to produce a uniform distribution of \\( g_{\\theta}(e_{P_{i,m}}) \\) on NE-Ring. If different semantics unevenly distributed on NE-Ring, the resulting discrete semantic values will be overly monotonous and the green list is more changeless. Consequently, the green list might be revealed by counting the token frequency, which compromises the concealment of watermark and leads to the risk of being cracked. Ideally, SemaMark should generate a wider variety of semantic values for different sentences, while each semantic value is robust and stable if its corresponding sentence is paraphrased. To achieve this goal, we propose to use Contrastive Learning to train MLP since Contrastive Learning has the property of uniformity that the data will be evenly distributed in the whole feature space (Wang and Isola, 2020). The uniform distribution can help the normalized vectors cover all the semantic values. As a result, NE-Ring can generate a wider variety of semantic values to prevent the watermark from being cracked."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.486,
+                0.929,
+                0.517,
+                0.941
+            ],
+            "angle": 0,
+            "content": "616"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.49,
+                0.198
+            ],
+            "angle": 0,
+            "content": "In Contrastive Learning, we first input the sentences into the model \\( f \\) to get a batch of sequences of \\( m \\) tokens and their pooling embeddings \\( e_{P_{i,m}} \\), denoted as \\( \\{e_j\\} \\), where \\( j \\in [B] \\) and \\( B \\) is the batch size. To compose a contrastive loss, we construct the positive and negative pairs by a soft augmentation:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.216,
+                0.21,
+                0.386,
+                0.23
+            ],
+            "angle": 0,
+            "content": "\\[\n\\boldsymbol {e} _ {j + B} = \\boldsymbol {e} _ {j} ^ {+} = \\boldsymbol {e} _ {j} + \\epsilon ,\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.241,
+                0.49,
+                0.402
+            ],
+            "angle": 0,
+            "content": "where \\(\\epsilon \\sim \\mathcal{N}(0, \\sigma^2)\\) is a Gaussian noise. The soft augmentation can simplify the construction of positive samples. With this soft augmentation, we can assign the samples sharing similar embeddings from the same sequence as positive pairs and samples from different sequences as negative pairs. This is consistent with our intuition that the paraphrased semantic embeddings will not change significantly and can remain robust. Then the contrastive loss is"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.121,
+                0.411,
+                0.482,
+                0.445
+            ],
+            "angle": 0,
+            "content": "\\[\nL _ {j} = - \\log \\frac {\\exp \\left(\\sin \\left(g _ {\\theta} (\\boldsymbol {e} _ {j}) , g _ {\\theta} (\\boldsymbol {e} _ {j} ^ {+})\\right) / \\tau\\right)}{\\sum_ {k \\neq j , k \\in [ 2 B ]} \\exp \\left(\\sin \\left(g _ {\\theta} (\\boldsymbol {e} _ {j}) , g _ {\\theta} (\\boldsymbol {e} _ {k})\\right) / \\tau\\right)},\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.455,
+                0.49,
+                0.55
+            ],
+            "angle": 0,
+            "content": "where \\(\\mathrm{sim}(\\cdot)\\) is cosine similarity and \\(\\tau\\) is the temperature. By Contrastive Learning, the output of reduced semantic embeddings can be evenly distributed in all of the space on NE-Ring, and cover all the discrete sections to improve the robustness of SemaMark."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.562,
+                0.303,
+                0.578
+            ],
+            "angle": 0,
+            "content": "3.3 \\(Q\\)-offset detection"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.13,
+                0.595,
+                0.475,
+                0.801
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.13,
+                0.81,
+                0.472,
+                0.826
+            ],
+            "angle": 0,
+            "content": "Figure 2: \\(Q\\)-offset detection vs. existing detection"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.842,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Existing detection methods check the matching between partition seed and the consequent tokens in a one-to-one manner as shown in Figure 2(a). The detection method first recalculates the seed for each token position and gets the partition of the green"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.885,
+                0.343
+            ],
+            "angle": 0,
+            "content": "list, and then checks whether the consequent token is in the partitioned green list token by token. In SemaMark, this strategy can be effective when the text is not paraphrased. However, after paraphrase, this detection could be suboptimal because the semantic values of some sequences which are close to the boundaries of the discrete section \\([i,i + 1)\\) might change as shown in Figure 2(b). This is because the window of \\(m\\) tokens will slide token by token during the auto-regressive generation process, and the semantic change will also accumulate when the window is sliding. The semantic values closed to the boundary usually happen when the change accumulates to some extent. This change of boundary semantic values will lead to some mismatch and reduce the accuracy like \\(\\tilde{t}_5\\) in Figure 2(b)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.349,
+                0.885,
+                0.655
+            ],
+            "angle": 0,
+            "content": "To mitigate the influence of this error, we propose \\(Q\\)-offset detection. As shown in Figure 2(c), we offset the discrete seed by \\(q\\) tokens to detect the matching between semantics and the consequent tokens, where \\(q \\in \\{-Q, -(Q - 1), \\dots, 0, 1, \\dots, Q\\}\\) and the sign of \\(q\\) indicates the direction of the offset. We choose the maximal \\(z\\)-statistic in different \\(q\\) as the \\(Q\\)-offset score. However, \\(Q\\)-offset detection will also increase the \\(Q\\)-offset score of non-watermark text, which indicates that the detected green word fraction \\(\\gamma\\) of non-watermark text is higher. The \\(\\gamma\\) in Eq. (1) is possibly inaccurate. Thus during generation, we set \\(\\gamma\\) to a fixed value, while in detection process, we treat \\(\\gamma\\) as a hyperparameter and use a validation set to determine its value in practice. In Section 4.4, we discuss the ablation studies of \\(Q\\)-offset and \\(\\gamma\\) and show that \\(Q\\)-offset can impressively improve the detection performance with robustness."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.688,
+                0.649,
+                0.705
+            ],
+            "angle": 0,
+            "content": "4 Experiment"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.729,
+                0.885,
+                0.922
+            ],
+            "angle": 0,
+            "content": "In this section, we conduct experiments to demonstrate the robustness of SemaMark. In Section 4.2, we demonstrate that its robustness is better than the baseline methods. In Section 4.3, we show that our watermark has almost no influence on the quality of generated texts. In Section 4.4, we use ablation studies to demonstrate the effectiveness of partition function and \\( Q \\)-offset detection, and show the sensitivity of the partition function. In Section 4.5 we visualize the distribution of NE-Ring and provide analysis on the feature distribution of Contractive Learning."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.486,
+                0.929,
+                0.516,
+                0.941
+            ],
+            "angle": 0,
+            "content": "617"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.132,
+                0.082,
+                0.868,
+                0.219
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\"></td><td rowspan=\"2\">Paraphrase</td><td colspan=\"4\">ROC-AUC</td><td colspan=\"4\">F1 with best threshold</td></tr><tr><td>LeftHash</td><td>SelfHash</td><td>EXP-Edit</td><td>ours</td><td>LeftHash</td><td>SelfHash</td><td>EXP-Edit</td><td>ours</td></tr><tr><td rowspan=\"4\">OPT-2.7B</td><td>No paraphrase</td><td>0.9913</td><td>0.9886</td><td>0.9799</td><td>0.9948</td><td>0.9921</td><td>0.9861</td><td>0.9708</td><td>0.9905</td></tr><tr><td>Translation</td><td>0.9091</td><td>0.8147</td><td>0.8749</td><td>0.9692</td><td>0.8456</td><td>0.7622</td><td>0.8157</td><td>0.9330</td></tr><tr><td>Dipper</td><td>0.9878</td><td>0.9728</td><td>0.9736</td><td>0.9911</td><td>0.9727</td><td>0.9400</td><td>0.9620</td><td>0.9701</td></tr><tr><td>GPT-3.5</td><td>0.9028</td><td>0.7908</td><td>0.9392</td><td>0.9406</td><td>0.8358</td><td>0.7378</td><td>0.8852</td><td>0.8902</td></tr><tr><td rowspan=\"4\">OPT-6.7B</td><td>No paraphrase</td><td>0.9918</td><td>0.9930</td><td>0.9784</td><td>0.9949</td><td>0.9911</td><td>0.9863</td><td>0.9705</td><td>0.9858</td></tr><tr><td>Translation</td><td>0.8807</td><td>0.8098</td><td>0.8625</td><td>0.9308</td><td>0.8129</td><td>0.7468</td><td>0.8013</td><td>0.8882</td></tr><tr><td>Dipper</td><td>0.9904</td><td>0.9747</td><td>0.9728</td><td>0.9871</td><td>0.9786</td><td>0.9432</td><td>0.9620</td><td>0.9821</td></tr><tr><td>GPT-3.5</td><td>0.8990</td><td>0.7909</td><td>0.8996</td><td>0.9377</td><td>0.8300</td><td>0.7367</td><td>0.8354</td><td>0.8766</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.114,
+                0.228,
+                0.882,
+                0.243
+            ],
+            "angle": 0,
+            "content": "Table 1: Watermark detection results under three paraphrases. (The best performance under paraphrase is bolded.)"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.268,
+                0.308,
+                0.284
+            ],
+            "angle": 0,
+            "content": "4.1 Experiment setups"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.291,
+                0.489,
+                0.466
+            ],
+            "angle": 0,
+            "content": "Backbone models and datasets. We test our method by watermarking two models, OPT-2.7B and OPT-6.7B (Zhang et al., 2022) which are referred to as the backbone models in following sections. For dataset, we use the news-like subset of C4 (Raffel et al., 2020), which covers a variety of topics. From the news-like subset of C4, we extract a training set, a validation set and a test set. For each sample, we use the first half of text as prompt to generate watermark sentences. More details can be found in Appendix A."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.469,
+                0.489,
+                0.629
+            ],
+            "angle": 0,
+            "content": "Baseline methods. We compare our method with three baselines LeftHash, SelfHash (Kirchenbauer et al., 2023b) and EXP-Edit (Kuditipudi et al., 2023). LeftHash and SelfHash are two methods based on the partition of vocabulary using the hashes of tokens. EXP-Edit uses a private sequence to encode the watermark by changing the probability distribution of the sequence of tokens. More details on the implementation can be found in Appendix A."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.632,
+                0.489,
+                0.856
+            ],
+            "angle": 0,
+            "content": "Paraphrase setups. We use three representative methods to paraphrase the watermarked text, round-trip translation (Tiedemann and Thottingal, 2020), Dipper (Krishna et al., 2023) and GPT-3.5. For round-trip translation, we first translate from English to another language and then transform back to English, such that some words and expressions will be changed because the translation is not an one-to-one mapping. For Dipper, we follow the parameter setting in Kirchenbauer et al. (2023b). For GPT-3.5, we use the prompt in Kirchenbauer et al. (2023b) to query GPT-3.5 for paraphrase. The examples of the three paraphrases can be found at Appendix B."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.858,
+                0.49,
+                0.921
+            ],
+            "angle": 0,
+            "content": "Evaluation metrics and hyper-parameters. We use F1 score with best threshold and ROC-AUC to measure the performance of the watermark detection. All the metrics are calculated based on"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.268,
+                0.885,
+                0.412
+            ],
+            "angle": 0,
+            "content": "at least 500 watermarked samples and 500 non-watermark samples. The length of watermarked samples before paraphrase and non-watermark samples is \\(200 \\pm 25\\). In generation, we set \\(\\gamma = 1/4\\) for LeftHash, SelfHash and SemaMark. In detection, we set \\(\\gamma = 1/3\\) and \\(\\delta = 2\\) based on the validation set in Section 4.4(b). In SemaMark, we set \\(m = 15\\), \\(Q = 15\\), \\(K = 5\\) for OPT-2.7B and \\(K = 4\\) for OPT-6.7B."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.429,
+                0.661,
+                0.443
+            ],
+            "angle": 0,
+            "content": "4.2 Main Results"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.453,
+                0.884,
+                0.679
+            ],
+            "angle": 0,
+            "content": "In this subsection, we demonstrate the robustness of the proposed SemaMark under paraphrase by comparing it with three baseline methods on two backbone models. We first generate watermarked texts and use three paraphrase methods to remove the watermarks. The detection performance of both texts with and without paraphrase is reported in Table 1. As we can see, before paraphrase, all the watermarked methods have good detection performance. After paraphrase, SemaMark has the best detection performance most of the time across all the backbone models and all the paraphrase methods, which suggests that our method is more robust against paraphrase."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.681,
+                0.885,
+                0.922
+            ],
+            "angle": 0,
+            "content": "In detail, by round-trip translation, the paraphrase reduces the detection ability of baseline methods effectively, while the watermark of SemaMark is robust. Under round-trip translation, the best ROC-AUC of baselines is 0.9091 on OPT-2.7B and 0.8807 on OPT-6.7B, respectively. But ROC-AUC of SemaMark is 0.9692 and 0.9308, which is at least 0.05 higher than all the baseline methods. Similarly, under paraphrase of GPT-3.5, SemaMark is better than all the baselines. The best baseline performance under GPT-3.5 is 0.9392 in ROC-AUC on OPT-2.7B and 0.8990 in ROC-AUC on opt-6.7B, but SemaMark has higher AUC-ROC of 0.9406 and 0.9377. For Dipper, we note that all methods are robust to Dipper since it does not sig"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.486,
+                0.929,
+                0.516,
+                0.941
+            ],
+            "angle": 0,
+            "content": "618"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.49,
+                0.214
+            ],
+            "angle": 0,
+            "content": "nificantly reduce the detection performance. However, SemaMark is still one of the most robust. On OPT-2.7B, it performs best in ROC-AUC, while on OPT-6.7B, it has the best F1 score. From Table 1, the results show an obvious improvement of SemaMark in robustness. This implies that using semantics as the seed for the partition function is effective under paraphrase."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.228,
+                0.26,
+                0.244
+            ],
+            "angle": 0,
+            "content": "4.3 Text Quality"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.119,
+                0.261,
+                0.3,
+                0.392
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.167,
+                0.398,
+                0.251,
+                0.41
+            ],
+            "angle": 0,
+            "content": "(a) OPT-2.7B"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.302,
+                0.262,
+                0.486,
+                0.393
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.351,
+                0.398,
+                0.435,
+                0.41
+            ],
+            "angle": 0,
+            "content": "(b) OPT-6.7B"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.184,
+                0.422,
+                0.419,
+                0.437
+            ],
+            "angle": 0,
+            "content": "Figure 3: Text quality (perplexity)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.455,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Watermark should not compromise the generation quality of LLMs. In this subsection, we compare the text quality by calculating perplexity and demonstrate that our watermark has almost no influence on the generated quality. Perplexity measures the likelihood that a sentence is generated by one model. Lower perplexity means the watermarked text is more predictable. In other words, it is more consistent with the reasoning of the given model. In Figure 3, we use OPT-6.7B with no watermark to get perplexity for all the watermarked methods. All the results in Figure 3 are calculated without paraphrase, because the generation quality of text is not related to paraphrase. From Figure 3a on OPT-2.7B, we can see that our watermark, LeftHash and SelfHash have almost no influence on the generation quality. They has perplexity at around 6 which is similar as the generated text without watermark. Instead, EXP-Edit has much higher perplexity, which means that EXP-Edit changes the generated text in an aggressive way and much reduces the generation quality after watermarking. This is probably because EXP-Edit adjusts the logits on the whole vocabulary. From Figure 3b, we can draw similar conclusions for OPT-6.7B. EXP-Exit also increases the perplexity by around 10, while the average perplexity of LeftHash, SelfHash and ours is around 1 higher than the non-watermarked generated text. In summary, our SemaMark can"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.56,
+                0.082,
+                0.81,
+                0.212
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.6,
+                0.22,
+                0.794,
+                0.234
+            ],
+            "angle": 0,
+            "content": "Figure 4: ROC-AUC and \\( m \\)"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.512,
+                0.246,
+                0.685,
+                0.337
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.521,
+                0.342,
+                0.69,
+                0.356
+            ],
+            "angle": 0,
+            "content": "(a) ROC-AUC and offset \\(Q\\)"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.697,
+                0.248,
+                0.865,
+                0.338
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.725,
+                0.342,
+                0.853,
+                0.356
+            ],
+            "angle": 0,
+            "content": "(b) ROC-AUC and \\(\\gamma\\)"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.578,
+                0.366,
+                0.813,
+                0.382
+            ],
+            "angle": 0,
+            "content": "Figure 5: Text quality (perplexity)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.407,
+                0.863,
+                0.423
+            ],
+            "angle": 0,
+            "content": "keep the quality and robustness simultaneously."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.433,
+                0.674,
+                0.449
+            ],
+            "angle": 0,
+            "content": "4.4 Ablation Study"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.454,
+                0.882,
+                0.517
+            ],
+            "angle": 0,
+            "content": "In this subsection, we study the influence of the length of the sequence we use for generating one semantic value and the sensitivity of the partition function."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.528,
+                0.885,
+                0.866
+            ],
+            "angle": 0,
+            "content": "a) Length of previous sequence tokens, \\( m \\). In the first step of SemaMark, i.e., weighted embedding pooling, we use the semantic of the previous \\( m \\) tokens to get the more stable embedding. But if the length of the sequence is too long, it will also hurt the robustness. In Figure 4, we test watermark on OPT-2.7B with different \\( m \\) and draw the ROC-AUC. The results show that before \\( m = 15 \\), ROC-AUC is in the trend of increase as the \\( m \\) changes. But when \\( m > 15 \\), ROC-AUC becomes fluctuating. It is possibly because that the distant tokens will include more change after paraphrase as we mentioned in Section 3.1. Another possible reason is that in the beginning of generation for the first \\( m \\) tokens, the number of previous tokens is smaller than \\( m \\) and NE-Ring can only use the embeddings of limited tokens for prediction, which may be unstable. Thus, too long or too short sequence will hurt the robustness of SemaMark against paraphrase. In our experiments, we choose \\( m = 15 \\) for all the settings."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.874,
+                0.884,
+                0.922
+            ],
+            "angle": 0,
+            "content": "b) \\(Q\\)-offset detection In this subsection, we show that the effectiveness of the proposed \\(Q\\)-offset detection. In Figure 5a, we demonstrate the change"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.486,
+                0.929,
+                0.516,
+                0.941
+            ],
+            "angle": 0,
+            "content": "619"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.49,
+                0.438
+            ],
+            "angle": 0,
+            "content": "of ROC-AUC of SemaMark with different \\( Q \\) in offset detection under three different paraphrases. \\( Q \\)-offset detection searches the highest \\( z \\)-statistics from \\( -Q \\) to \\( Q \\) as the \\( Q \\)-offset score. From Figure 5a, we can see that when \\( Q \\) increases, ROC-AUC first increases and decreases after \\( Q \\) is around 15. When \\( Q < 15 \\), the offset can help correct the errors of semantic values close to the boundary. Compared with detection without offset, i.e. \\( Q = 0 \\), ROC-AUC of SemaMark is much better, which means that the offset can help to solve the errors of semantic values around the boundaries that are more vulnerable to paraphrases. When \\( Q > 15 \\), the correction of this error is limited, because the offset will also increase the \\( Q \\)-offset score of negative samples as it also searches the highest \\( z \\)-statistics of negative samples. On the other hand, the computation cost will also increase if \\( Q \\) is too large because it has to search more possible \\( q \\). In practice, we set \\( Q = 15 \\) in all the experiments, which can effectively reduce the influence of the errors of semantic values at the boundaries."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.44,
+                0.49,
+                0.634
+            ],
+            "angle": 0,
+            "content": "Since the \\(Q\\)-offset detection searches the highest green word fraction, the fraction of green list word of non-watermarked text will be higher than the \\(\\gamma\\) that we used to randomly select the green list. Thus, it is not accurate to use the original \\(\\gamma\\) for \\(z\\)-statistics. We treat \\(\\gamma\\) as a hyper-parameter and use a validation set to select its value. As shown in Figure 5b, the detection performance of SemaMark under paraphrases of Dipper and GPT-3.5 will reach the highest when \\(\\gamma\\) is around \\(1/3\\), while it will continue to increase under paraphrase. In practice, we set \\(\\gamma = 1/3\\) for \\(Q\\)-offset detection."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.648,
+                0.492,
+                0.922
+            ],
+            "angle": 0,
+            "content": "c) Sensitivity of partition function. As we mentioned, the partition function is sensitive to any change of the input because it only uses the input as the seed of the random generator. To validate its sensitivity to continuous embeddings, we adopt the embedding vector as the input to show that, with tiny change of the embeddings, the partition of vocabulary can be very different. We propose a hash method based on md5sum (Deepakumara et al., 2001) to adopt the partition function by transforming the continuous embeddings to an integral seed. We use 1000 sequences to test the sensitivity. For each sequence embedding, we first get a green list from the partition function. Then we change one dimension of the embedding by only 1e-5 to get a new partition result. The overlapping of the green list before and after changing is \\(24.99\\%\\) on the av"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.514,
+                0.082,
+                0.881,
+                0.126
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\"></td><td colspan=\"3\">ROC-AUC</td><td colspan=\"3\">F1 with best threshold</td></tr><tr><td>LeftHash</td><td>SelfHash</td><td>ours</td><td>LeftHash</td><td>SelfHash</td><td>ours</td></tr><tr><td>LLaMA-7B</td><td>0.819</td><td>0.838</td><td>0.846</td><td>0.748</td><td>0.774</td><td>0.781</td></tr><tr><td>LLaMA2-7B</td><td>0.811</td><td>0.841</td><td>0.872</td><td>0.749</td><td>0.773</td><td>0.810</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.509,
+                0.135,
+                0.883,
+                0.163
+            ],
+            "angle": 0,
+            "content": "Table 2: Watermark detection results under different model size.)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.191,
+                0.885,
+                0.415
+            ],
+            "angle": 0,
+            "content": "erage of 1000 sequences. It is consistent with \\(\\gamma\\) we use to watermark, because the random partition with the changed embedding is independent from the original one. It means the partition function is sensitive to any small change in its input. Instead, after we use NE-Ring to discretize the embeddings, the overlapping of green list after changing embeddings by 1e-5 is \\(100\\%\\), which means the discretization can effectively handle this change. In practice, SemaMark can provide the tolerance that is much larger than 1e-5, which makes the watermark more robust under paraphrase. With the improvement of \\(Q\\)-offset, the detection of SemaMark is more robust and effective."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.428,
+                0.887,
+                0.636
+            ],
+            "angle": 0,
+            "content": "\\(d\\) ) Model size. To show the robustness of our method on different model sizes, in this section, we also test the watermark under round-trip translation paraphrase on LLaMA-7B and LLaMA2-7B, which have larger size and different architectures. As indicated in Table 3, our approach consistently exhibits the highest robustness against paraphrasing. Specifically, in the LLaMA2-7B model, SemaMark significantly outperforms the baseline models, achieving an increase of 0.06 and 0.03 in ROC-AUC. Similarly, in the LLaMA-7B model, our method shows superior performance with an increase of 0.027 and 0.009 in ROC-AUC."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.65,
+                0.855,
+                0.667
+            ],
+            "angle": 0,
+            "content": "4.5 Distribution on NE-Ring based on CL"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.517,
+                0.683,
+                0.691,
+                0.781
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.568,
+                0.785,
+                0.646,
+                0.797
+            ],
+            "angle": 0,
+            "content": "(a) NE-Ring"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.697,
+                0.684,
+                0.876,
+                0.78
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.722,
+                0.785,
+                0.851,
+                0.797
+            ],
+            "angle": 0,
+            "content": "(b) Distribution on \\(\\phi\\)"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.575,
+                0.809,
+                0.819,
+                0.824
+            ],
+            "angle": 0,
+            "content": "Figure 6: Visualization of NE-Ring"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.842,
+                0.885,
+                0.922
+            ],
+            "angle": 0,
+            "content": "In this subsection, we demonstrate that Contrastive Learning can help evenly distribute the semantics on the NE-Ring. The even distribution can help the sequences reach all possible semantic values and provide more diverse semantic values"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.486,
+                0.929,
+                0.517,
+                0.941
+            ],
+            "angle": 0,
+            "content": "620"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.49,
+                0.376
+            ],
+            "angle": 0,
+            "content": "to prevent the watermark from being cracked by counting token frequency. In Figure 6a, we use Gaussian density estimation (Chen, 2017) to get the distribution of the semantics on the NE-Ring before discretization. We use different colors to show the density. The NE-Ring in Figure 6a shows that, the distribution is uniform. All the density is between 0.052 and 0.054. We further plot the density based on the polar angle \\(\\phi\\) in Figure 6b where the density has almost no change on all the polar angle from 0 to \\(2\\pi\\). This implies that the training based on Contrastive Learning can ensure the semantics will reach all possible discrete values. It can prevent the case where the discrete values will gather in some discrete sections and produce monotonous vocabulary partitions. As a result, it can protect the watermark from being cracked by counting token frequency."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.388,
+                0.248,
+                0.403
+            ],
+            "angle": 0,
+            "content": "5 Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.415,
+                0.49,
+                0.721
+            ],
+            "angle": 0,
+            "content": "In this paper, we use the semantic information for watermarking to enhance the robustness against paraphrase. The existing watermark methods use the matching between the previous tokens and the partition vocabulary. This matching can be easily broken by paraphrase. However, we construct the mapping between the semantics and the vocabulary. In this way, the semantics will stay stable under paraphrase and the robustness of watermark can be enhanced. To make use of semantics, we propose SemaMark to discretize the embedding space on NE-Ring and propose a training method based on CL. In addition, we use \\(Q\\)-offset detection to further advance the robustness by increasing the tolerance of the semantic values close to the discrete boundary. In experiments, we demonstrate our method can perform much better compared with baseline methods under paraphrase while having little influence on the generation quality."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.735,
+                0.251,
+                0.75
+            ],
+            "angle": 0,
+            "content": "6 Limitations"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.762,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": "In some cases, the customers may rely on some API-based LLMs and do not have the access to the embeddings and the permission to modify the logits during generation. Although our watermark method can effectively detect the LLM-generated content and increase the detection success rate under paraphrase, it is not applicable for black-box LLMs. The second weakness of our method is that the NE-Ring is dependent on the semantic embedding of LLMs. For each LLM, we need to train a"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.885,
+                0.213
+            ],
+            "angle": 0,
+            "content": "specialized EN-Ring, which might be inflexible if we want to produce a general model for NE-Ring or fine-tune the LLMs. Despite the weaknesses, our method is successful in the problem of robustness under paraphrase. In the future work, we will continue to extent our method into black-box LLMs and a universal model that does not require customized training for various specific LLMs."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.214,
+                0.886,
+                0.342
+            ],
+            "angle": 0,
+            "content": "Potential risk. Our discussion about the robustness might provide motivation for the attackers to find other methods like adaptive attack. Although we provide robustness under paraphrase, if the unauthorized people propose possible attack method focusing on the green-list based watermark from other perspectives, the detection rate for LLM-generated texts are still possible to be reduced."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.354,
+                0.714,
+                0.371
+            ],
+            "angle": 0,
+            "content": "7 Acknowledgements"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.38,
+                0.885,
+                0.54
+            ],
+            "angle": 0,
+            "content": "Jie Ren, Han Xu, Yingqian Cui, and Jiliang Tang are supported by the National Science Foundation (NSF) under grant numbers CNS 2246050, IIS1845081, IIS2212032, IIS2212144, IOS2107215, DUE 2234015, DRL 2025244 and IOS2035472, the Army Research Office (ARO) under grant number W911NF-21-1-0198, the Home Depot, Cisco Systems Inc, Amazon Faculty Award, Johnson&Johnson, JP Morgan Faculty Award and SNAP."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.567,
+                0.61,
+                0.582
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.51,
+                0.59,
+                0.885,
+                0.643
+            ],
+            "angle": 0,
+            "content": "Steven Bird, Ewan Klein, and Edward Loper. 2009. Natural language processing with Python: analyzing text with the natural language toolkit. \"O'Reilly Media, Inc.\""
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.653,
+                0.885,
+                0.719
+            ],
+            "angle": 0,
+            "content": "Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.729,
+                0.887,
+                0.77
+            ],
+            "angle": 0,
+            "content": "Yen-Chi Chen. 2017. A tutorial on kernel density estimation and recent advances. Biostatistics & Epidemiology, 1(1):161-187."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.779,
+                0.885,
+                0.833
+            ],
+            "angle": 0,
+            "content": "Antonia Creswell, Murray Shanahan, and Irina Higgins. 2022. Selection-inference: Exploiting large language models for interpretable logical reasoning. arXiv preprint arXiv:2205.09712."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.51,
+                0.842,
+                0.885,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Janaka Deepakumara, Howard M Heys, and R Venkatesan. 2001. FPGA implementation of md5 hash algorithm. In Canadian Conference on Electrical and Computer Engineering 2001. Conference Proceedings (Cat. No. 01TH8555), volume 2, pages 919-924. IEEE."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.51,
+                0.59,
+                0.887,
+                0.92
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.486,
+                0.929,
+                0.514,
+                0.941
+            ],
+            "angle": 0,
+            "content": "621"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.166
+            ],
+            "angle": 0,
+            "content": "Tianyu Gao, Xingcheng Yao, and Danqi Chen. 2021. Simcse: Simple contrastive learning of sentence embeddings. In 2021 Conference on Empirical Methods in Natural Language Processing, EMNLP 2021, pages 6894-6910. Association for Computational Linguistics (ACL)."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.175,
+                0.49,
+                0.267
+            ],
+            "angle": 0,
+            "content": "John Giorgi, Osvald Nitski, Bo Wang, and Gary Bader. 2021. Declutr: Deep contrastive learning for unsupervised textual representations. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 879-895."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.276,
+                0.488,
+                0.342
+            ],
+            "angle": 0,
+            "content": "Biyang Guo, Xin Zhang, Ziyuan Wang, Minqi Jiang, Jinran Nie, Yuxuan Ding, Jianwei Yue, and Yupeng Wu. 2023. How close is chatgpt to human experts? comparison corpus, evaluation, and detection. arXiv preprint arXiv:2301.07597."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.351,
+                0.488,
+                0.494
+            ],
+            "angle": 0,
+            "content": "Charles R. Harris, K. Jarrod Millman, Stefan J. van der Walt, Ralf Gommers, Pauli Virtanen, David Cournaepau, Eric Wieser, Julian Taylor, Sebastian Berg, Nathaniel J. Smith, Robert Kern, Matti Picus, Stephan Hoyer, Marten H. van Kerkwijk, Matthew Brett, Allan Haldane, Jaime Fernandez del Rio, Mark Wiebe, Pearu Peterson, Pierre Gerard-Marchant, Kevin Sheppard, Tyler Reddy, Warren Weckesser, Hameer Abbasi, Christoph Gohlke, and Travis E. Oliphant. 2020. Array programming with NumPy. Nature, 585(7825):357-362."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.505,
+                0.488,
+                0.557
+            ],
+            "angle": 0,
+            "content": "Jean Kaddour, Joshua Harris, Maximilian Mozes, Herb Bradley, Roberta Raileanu, and Robert McHardy. 2023. Challenges and applications of large language models. arXiv preprint arXiv:2307.10169."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.567,
+                0.488,
+                0.62
+            ],
+            "angle": 0,
+            "content": "John Kirchenbauer, Jonas Geiping, Yuxin Wen, Jonathan Katz, Ian Miers, and Tom Goldstein. 2023a. A watermark for large language models. arXiv preprint arXiv:2301.10226."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.629,
+                0.488,
+                0.707
+            ],
+            "angle": 0,
+            "content": "John Kirchenbauer, Jonas Geiping, Yuxin Wen, Manli Shu, Khalid Saifullah, Kezhi Kong, Kasun Fernando, Aniruddha Saha, Micah Goldblum, and Tom Goldstein. 2023b. On the reliability of watermarks for large language models. arXiv preprint arXiv:2306.04634."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.717,
+                0.488,
+                0.782
+            ],
+            "angle": 0,
+            "content": "Sarah Kreps, R Miles McCain, and Miles Brundage. 2022. All the news that's fit to fabricate: A-generated text as a tool of media misinformation. Journal of experimental political science, 9(1):104-117."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.793,
+                0.488,
+                0.857
+            ],
+            "angle": 0,
+            "content": "Kalpesh Krishna, Yixiao Song, Marzena Karpinska, John Wieting, and Mohit Iyyer. 2023. Paraphrasing evades detectors of ai-generated text, but retrieval is an effective defense. arXiv preprint arXiv:2303.13408."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.868,
+                0.488,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Rohith Kuditipudi, John Thickstun, Tatsunori Hashimoto, and Percy Liang. 2023. Robust distortion-free watermarks for language models. arXiv preprint arXiv:2307.15593."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.086,
+                0.882,
+                0.166
+            ],
+            "angle": 0,
+            "content": "Bohan Li, Hao Zhou, Junxian He, Mingxuan Wang, Yiming Yang, and Lei Li. 2020. On the sentence embeddings from pre-trained language models. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 9119-9130."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.178,
+                0.882,
+                0.217
+            ],
+            "angle": 0,
+            "content": "Linyang Li, Pengyu Wang, Ke Ren, Tianxiang Sun, and Xipeng Qiu. 2023. Origin tracing and detecting of llms. arXiv preprint arXiv:2304.14072."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.23,
+                0.882,
+                0.281
+            ],
+            "angle": 0,
+            "content": "Aiwei Liu, Leyi Pan, Xuming Hu, Shiao Meng, and Lijie Wen. 2023a. A semantic invariant robust watermark for large language models. arXiv preprint arXiv:2310.06356."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.295,
+                0.882,
+                0.347
+            ],
+            "angle": 0,
+            "content": "Aiwei Liu, Leyi Pan, Yijian Lu, Jingjing Li, Xuming Hu, Lijie Wen, Irwin King, and Philip S Yu. 2023b. A survey of text watermarking in the era of large language models. arXiv preprint arXiv:2312.07913."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.36,
+                0.882,
+                0.439
+            ],
+            "angle": 0,
+            "content": "Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. 2022. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.451,
+                0.882,
+                0.516
+            ],
+            "angle": 0,
+            "content": "Eric Mitchell, Yoonho Lee, Alexander Khazatsky, Christopher D Manning, and Chelsea Finn. 2023. Detectgpt: Zero-shot machine-generated text detection using probability curvature. arXiv preprint arXiv:2301.11305."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.529,
+                0.882,
+                0.607
+            ],
+            "angle": 0,
+            "content": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.621,
+                0.882,
+                0.699
+            ],
+            "angle": 0,
+            "content": "Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3982-3992."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.712,
+                0.882,
+                0.777
+            ],
+            "angle": 0,
+            "content": "Jie Ren, Han Xu, Pengfei He, Yingqian Cui, Shenglai Zeng, Jiankun Zhang, Hongzhi Wen, Jiayuan Ding, Hui Liu, Yi Chang, et al. 2024. Copyright protection in generative ai: A technical perspective. arXiv preprint arXiv:2402.02333."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.79,
+                0.882,
+                0.843
+            ],
+            "angle": 0,
+            "content": "Vinu Sankar Sadasivan, Aounon Kumar, Sriram Balasubramanian, Wenxiao Wang, and Soheil Feizi. 2023. Can ai-generated text be reliably detected? arXiv preprint arXiv:2303.11156."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.855,
+                0.882,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Joni Salminen, Chandrashekhar Kandpal, Ahmed Mohamed Kamel, Soon-gyo Jung, and Bernard J Jansen. 2022. Creating and detecting fake reviews of online products. Journal of Retailing and Consumer Services, 64:102771."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.512,
+                0.086,
+                0.882,
+                0.92
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.486,
+                0.929,
+                0.516,
+                0.941
+            ],
+            "angle": 0,
+            "content": "622"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.086,
+                0.487,
+                0.126
+            ],
+            "angle": 0,
+            "content": "Ruixiang Tang, Yu-Neng Chuang, and Xia Hu. 2023. The science of detecting llm-generated texts. arXiv preprint arXiv:2303.07205."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.137,
+                0.487,
+                0.203
+            ],
+            "angle": 0,
+            "content": "Jörg Tiedemann and Santhosh Thottingal. 2020. OPUSMT — Building open translation services for the World. In Proceedings of the 22nd Annual Conference of the European Association for Machine Translation (EAMT), Lisbon, Portugal."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.215,
+                0.486,
+                0.253
+            ],
+            "angle": 0,
+            "content": "Saranya Venkatraman, Adaku Uchendu, and Dongwon Lee. 2023. Gpt-who: An information density-based machine-generated text detector."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.266,
+                0.487,
+                0.304
+            ],
+            "angle": 0,
+            "content": "Pengyu Wang, Linyang Li, Ke Ren, Botian Jiang, Dong Zhang, and Xipeng Qiu. 2023. Seqxgpt: Sentence-level ai-generated text detection."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.316,
+                0.487,
+                0.38
+            ],
+            "angle": 0,
+            "content": "Tongzhou Wang and Phillip Isola. 2020. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International Conference on Machine Learning, pages 9929-9939. PMLR."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.393,
+                0.487,
+                0.458
+            ],
+            "angle": 0,
+            "content": "Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35:24824-24837."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.47,
+                0.487,
+                0.534
+            ],
+            "angle": 0,
+            "content": "Junchao Wu, Shu Yang, Runzhe Zhan, Yulin Yuan, Derek F Wong, and Lidia S Chao. 2023. A survey on llm-gernerated text detection: Necessity, methods, and future directions. arXiv preprint arXiv:2310.14724."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.547,
+                0.487,
+                0.613
+            ],
+            "angle": 0,
+            "content": "Frank F Xu, Uri Alon, Graham Neubig, and Vincent Josua Hellendoorn. 2022. A systematic evaluation of large language models of code. In Proceedings of the 6th ACM SIGPLAN International Symposium on Machine Programming, pages 1-10."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.624,
+                0.486,
+                0.676
+            ],
+            "angle": 0,
+            "content": "Han Xu, Jie Ren, Pengfei He, Shenglai Zeng, Yingqian Cui, Amy Liu, Hui Liu, and Jiliang Tang. 2023. On the generalization of training-based chatgpt detection methods. arXiv preprint arXiv:2310.01307."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.688,
+                0.487,
+                0.766
+            ],
+            "angle": 0,
+            "content": "KiYoon Yoo, Wonhyuk Ahn, Jiho Jang, and Nojun Kwak. 2023. Robust multi-bit natural language watermarking through invariant features. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2092-2115."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.778,
+                0.487,
+                0.843
+            ],
+            "angle": 0,
+            "content": "Rowan Zellers, Ari Holtzman, Hannah Rashkin, Yonatan Bisk, Ali Farhadi, Franziska Roesner, and Yejin Choi. 2019. Defending against neural fake news. Advances in neural information processing systems, 32."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.855,
+                0.487,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. 2022. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.086,
+                0.487,
+                0.92
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.086,
+                0.882,
+                0.138
+            ],
+            "angle": 0,
+            "content": "Xuandong Zhao, Prabhanjan Ananth, Lei Li, and Yu-Xiang Wang. 2023. Provable robust watermarking for ai-generated text. arXiv preprint arXiv:2306.17439."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.486,
+                0.929,
+                0.516,
+                0.941
+            ],
+            "angle": 0,
+            "content": "623"
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.085,
+                0.475,
+                0.101
+            ],
+            "angle": 0,
+            "content": "A More details on experimental settings"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.11,
+                0.49,
+                0.302
+            ],
+            "angle": 0,
+            "content": "All the baseline models, backbone models and datasets we use are open source and available for academic purpose. For backbone models, we use the open-sourced model from Huggingface<sup>1</sup>. The implementation is based on Pytorch<sup>2</sup> framework and also depend on packages including NLTK (Bird et al., 2009) and Numpy (Harris et al., 2020). For baseline methods, we use the released official code from the authors. For paraphrase models, we use OPUS-MT translation model and Dipper on Huggingface repository<sup>3</sup>, and API of ChatGPT<sup>4</sup>."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.315,
+                0.371,
+                0.331
+            ],
+            "angle": 0,
+            "content": "B Examples of paraphrases"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.34,
+                0.49,
+                0.372
+            ],
+            "angle": 0,
+            "content": "In Table 3, we provide the examples of three paraphrases."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.138,
+                0.857,
+                0.395,
+                0.87
+            ],
+            "angle": 0,
+            "content": "\\(^{1}\\)https://huggingface.co/facebook/opt-2.7b"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.138,
+                0.87,
+                0.264,
+                0.883
+            ],
+            "angle": 0,
+            "content": "\\(^{2}\\)https://pytorch.org/"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.118,
+                0.883,
+                0.486,
+                0.907
+            ],
+            "angle": 0,
+            "content": "<sup>3</sup>https://huggingface.co/Helsinki-NLP/opus-mt-en-zh and https://huggingface.co/kalpeshk2011/dipper-paraphraser-xxl"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.138,
+                0.907,
+                0.317,
+                0.921
+            ],
+            "angle": 0,
+            "content": "<sup>4</sup>https://chat.openai.com/chat"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.486,
+                0.929,
+                0.516,
+                0.94
+            ],
+            "angle": 0,
+            "content": "624"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.139,
+                0.106,
+                0.86,
+                0.869
+            ],
+            "angle": 0,
+            "content": "<table><tr><td></td><td>Original</td><td>Paraphrased</td></tr><tr><td>Translation</td><td>The course ran from Feb. 16 to Feb. 18 and was designed to address officers&#x27; response to critical incidents and to enhance officer safety. &quot;I am excited by the training and professionalism demonstrated by the officers at the training session. They were receptive and focused on learning new techniques and approaches to policing.&quot; Com-missioner Walwyn added.</td><td>The training, which was held from 16 February to 18 February, was aimed at discussing police responses to major incidents and enhancing police security. &quot; I am excited about the training and professionalism shown by police officers in the course, who are willing to accept and focus on new policing techniques and methods. &quot; Com-missioner Walwyn adds.</td></tr><tr><td>GPT</td><td>She tells him she wants to protect him. She asks what else she can do. Stefan tells her not to trust Chad. She vows she won&#x27;t. Stefan offers to get rid of Chad for her if she wants. She says she wants Chad and his family gone. She wants to have his babies. She tells him to get out of her way. She doesn&#x27;t know if she can love him, but she wants his babies. She wants to be a mother. Stefan tells her she deserves a real man. He&#x27;ll love her and protect her. She says she wants him and his babies. She wants to have his babies. Stefan says she deserves a real man. He&#x27;ll love her and protect her. She says she wants him and his babies. She wants to have his babies. Stefan says she deserves a real man. He&#x27;ll love her and protect her. She says she wants him and his babies. She wants to have his babies. Stefan says she deserves a real man. He&#x27;ll love her and protect her. She says she wants him and his babies.</td><td>According to Reuters, China has already taken steps to solidify its presence in the South China Sea. It has constructed an airstrip on one island and In a tense conversation, a woman expresses her desire to protect a man, Stefan, and seeks his advice on how to do so. Stefan advises her not to trust a person named Chad, to which she firmly pledges her distrust. Stefan offers to eliminate Chad from her life if she desires, and she responds by expressing her wish for Chad and his family to disappear entirely. In a surprising revelation, she declares her desire to have Stefan&#x27;s children, emphasizing her longing to become a mother. Despite uncertain feelings of love towards Stefan, she insists on wanting his babies and urges him to step aside. Stefan, in turn, assures her that she deserves a genuine partner who will love and safeguard her. Yet, undeterred, she asserts her desire for both Stefan and his offspring, reiterat-ing her ambition to have his babies.</td></tr><tr><td>DIPPER</td><td>Kewanee got the ball again with 7.2 seconds to go, but couldn&#x27;t get a shot off. &quot;I think it was something we needed to have a plan for and it was one that worked in the end. We did it well enough,&quot; Keene said. &quot;We were trying to get to the rim. A lot of times when you are in a situation like that and you can&#x27;t score, it&#x27;s better to just foul.&quot; Macomb was 12-for-27 from behind the arc (42 percent). The Grinnell offense is centered around perimeter shooters. The two perimeter players for Macomb are Emerick and Reilly Rieder. The next closest shooter on the team is Rieder with 10 points. Nolan led Kewanee with 15 points. Macomb, now 14-5, now travels to Jacksonville on Thursday to play Jacksonville</td><td>Kewanee had the ball again with 7.2 seconds left but couldn&#x27;t get off a shot. &quot;I think that&#x27;s something we needed to have a plan for and in the end, it worked,&quot; coach Lambert said. &quot;We did it well enough,&quot; Keene said. &quot;We were trying to get to the rim. Often when you are in a situation like that and you can&#x27;t score, it&#x27;s better to foul.&quot; Macomb was 12 for 27 from beyond the arc (42 percent). The Grinnell offense is based on sharpshooting players. Macomb&#x27;s two shooters are Emerick and Rieder. Rieder has ten points. Nolan led Kewanee with 15 points. Macomb, now 14-5, will play at Jacksonville Thursday.</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.391,
+                0.877,
+                0.604,
+                0.892
+            ],
+            "angle": 0,
+            "content": "Table 3: Paraphrase examples."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.486,
+                0.929,
+                0.516,
+                0.941
+            ],
+            "angle": 0,
+            "content": "625"
+        }
+    ]
+]

2024/A Robust Semantics-based Watermark for Large Language Model against Paraphrasing/b1276659-b3bd-43b9-95dd-a7cb02ea87a5_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3e47c71a8919e9fe7fe81e8b811c2e7dc1a444612db1804a9951421a499dfe02
+size 1384238

2024/A Robust Semantics-based Watermark for Large Language Model against Paraphrasing/full.md

@@ -0,0 +1,267 @@
+# A Robust Semantics-based Watermark for Large Language Models against Paraphrasing
+
+Jie Ren $^{1}$ , Han Xu $^{1}$ , Yiding Liu $^{2}$ , Yingqian Cui $^{1}$ , Shuaiqiang Wang $^{2}$ , Dawei Yin $^{2}$ , Jiliang Tang $^{1}$
+
+<sup>1</sup>Michigan State University, <sup>2</sup>Baidu Inc.
+
+{renjie3, xuhan1, cuiyingq, tangjili}@msu.edu
+
+liuyiding.tanh@gmail.com, shqiang.wang@gmail.com, yindawei@acm.org
+
+# Abstract
+
+Large language models (LLMs) have show their remarkable ability in various natural language tasks. However, there are concerns that LLMs are possible to be used improperly or even illegally. To prevent the malicious usage of LLMs, detecting LLM-generated text becomes crucial in the deployment of LLM applications. Watermarking is an effective strategy to detect the LLM-generated content by encoding a pre-defined secret watermark to facilitate the detection process. However, the majority of existing watermark methods leverage the simple hashes of precedent tokens to partition vocabulary. Such watermarks can be easily eliminated by paraphrase and, correspondingly, the detection effectiveness will be greatly compromised. Thus, to enhance the robustness against paraphrase, we propose a semantics-based watermark framework, SemaMark. It leverages the semantics as an alternative to simple hashes of tokens since the semantic meaning of the sentences will be likely preserved under paraphrase and the watermark can remain robust. Comprehensive experiments are conducted to demonstrate the effectiveness and robustness of SemaMark under different paraphrases. Our code is available at github.com/renjie3/SemaMark.
+
+# 1 Introduction
+
+Large language models (LLMs) have shown their great ability in various natural language processing (NLP) tasks like Question Answering (QA) (Lu et al., 2022), reasoning tasks (Wei et al., 2022; Creswell et al., 2022) and code development (Xu et al., 2022). However, tremendous concerns have been raised that LLMs are possible to be used improperly and illegally. For example, indistinguishable fake news are easy to be fabricated (Kreps et al., 2022; Zellers et al., 2019) by language models, which, when disseminated, could instigate widespread panic. Similarly, in the commercial sphere, convincingly generated reviews
+
+can manipulate consumer perceptions, leading to unethical business competition (Salminen et al., 2022). Therefore, detecting LLM-generated text has become crucial in the real-world applications of LLMs (Wu et al., 2023; Sadasivan et al., 2023; Xu et al., 2023).
+
+Among diverse methods to detect LLM-generated texts, the watermark strategies have demonstrated outstanding precision (Liu et al., 2023b; Tang et al., 2023; Ren et al., 2024). It is proposed to encode a secret watermark into the generated texts, such that we can tell whether a text is generated by detecting this watermark. One representative strategy (Kirchenbauer et al., 2023a; Yoo et al., 2023) is to encode the watermark based on the "partition of vocabulary". In detail, given a language model, these methods devise a mapping from precedent tokens to a particular partition of the vocabulary by a partition function for the consequent token. The partition function leverages the hashes of the input as the seed of a random generator to split the vocabulary to a green list and a red list. During the text generation phase, the consequent token has an increased probability to be sampled from the green list. In this way, the watermark is encoded through the matching between the precedent tokens and the vocabulary partition for the consequent token. The detection is also facilitated by detecting this matching in generated contents. However, recent works (Krishna et al., 2023; Kaddour et al., 2023) reveal that this watermark may be easily eliminated by sentence paraphrasing. Individuals seeking to improperly utilize LLMs without being detected can paraphrase the generated contents, like altering the order and the choices of the words, and only retain the general meaning of the text to achieve their malicious goals like faking news. These paraphrases will change the seed of the partition function, i.e. the token hashes, and as we show in the Section 4.4, the partition function is sensitive to small changes. Consequently,
+
+the matching between the precedent tokens and the green list will be disrupted, and the detection effectiveness of the watermark can be dramatically compromised.
+
+In this paper, we propose to leverage the semantic meaning of precedent token sequences as the seed for partition function, instead of simple hashes of precedent tokens, since the core semantic meaning is expected to be maintained after paraphrase. To achieve this goal, one key obstacle is how to capture the semantics when applying them for the partition function to watermark the generated texts. It is a common practice to quantify the semantics via embeddings (Reimers and Gurevych, 2019; Gao et al., 2021; Li et al., 2020; Giorgi et al., 2021). Embeddings indeed can represent consistent semantics after paraphrase. Since the embeddings are high-dimensional vectors in the continuous space, they often present some minor changes after paraphrase. Although the main semantics are preserved, these minor changes can lead to a substantial difference in the partition of vocabulary because the random generator in the partition function is sensitive to the change of the seed, as shown in Section 4.4.
+
+To overcome the above challenge, i.e., to make the quantified semantics invariant and make the watermark robust under paraphrase, we propose a new watermark method, SemaMark, which discretizes the continuous embedding space. Intuitively, the discretization can coarsen the representation of the embeddings which could tolerate the potential minor changes caused by paraphrase. By proper discretization, the paraphrased semantics could stay in the same discrete section with a high probability and the discretized quantified semantics will likely remain the same even after paraphrase. Therefore, the partition results will not change. However, directly converting the high-dimensional embedding space into discrete is intricate and challenging. For example, discretizing each dimension will lead to a large amount of discrete values which is exponential to the number of dimensions. Thus, the minor changes by paraphrase can still cause the change of discrete values because the number of discrete values are too dense and each discrete value can tolerate only small changes. Therefore, the minor changes of high-dimensional embeddings can have a strong impact on the partition function. To address this problem, SemaMark first uses a MultiLayer Perception (MLP) to condense the continuous high-dimensional embeddings into normalized vectors in 2D space. The vectors are located
+
+on a unit circle named Normalized Embedding Ring (NE-Ring). Then the condensed NE-Ring is equally divided into various sections, transforming the continuous space into distinct discrete values, i.e., "semantic values". Based on the discretization, SemaMark further introduces two strategies to advance the watermark's concealment and to improve the robustness under paraphrase. First, SemaMark leverages the uniformity (Wang and Isola, 2020) of Contrastive Learning(CL) (Chen et al., 2020) to strength the MLP and mitigate the problem that the semantics are unevenly concentrating on some discrete sections on NE-Ring. The unevenly distribution will cause the final discrete semantic values overly monotonous. It raises the concern that the watermark might be cracked by counting token frequency (Zhao et al., 2023). Second, SemaMark utilizes an offset detection method to further enhance the robustness at the boundary of different discrete sections whose semantic values are possibly vulnerable to paraphrase. Comprehensive experiments are conducted to demonstrate the effectiveness and robustness of SemaMark under different paraphrases.
+
+# 2 Related works
+
+LLM-generated detection. As the development of LLMs, various LLM-generated detection tools have also been proposed. Learning-based methods train a classification model to detect the difference between human-written text and machine-generated text like Guo et al. (2023); Wang et al. (2023); Li et al. (2023). Other works do not rely on the classification model, but try to use the property of the LLM to test whether a given text is generated by LLMs. For example, DetectGPT (Mitchell et al., 2023) assumes that the generated text will have high likelihood. GPT-who (Venkatraman et al., 2023) uses UID-based features to model the unique statistical signature of each LLM and human author for accurate authorship attribution. These methods do not interact the generation process of LLMs and thus have to explore unknown features of LLMs for detection. Instead, watermarks can change the model with a small but pre-defined rule which accelerates the detection process effectively.
+
+Watermark. The distinction between watermark and other methods is that watermark can proactively change the generation to insert a concealed watermark into the generated text. This gives clear difference between watermarked and non
+
+![](images/fa529ed5c73dcbd8cd94fb9e4b2e9ab8a405affdc48e4ec8aaa035196bf24ffa.jpg)  
+Figure 1: The watermarking process of SemaMark
+
+watermarked texts. Watermark shifts the text using a small but pre-defined rule to make the detection much more effective. The partition of the vocabulary for each token is a representative watermark method (Kirchenbauer et al., 2023a; Yoo et al., 2023; Kirchenbauer et al., 2023b). In each autoregressive step of generating one token, the method uses the previous tokens' hashes, to select a part of the vocabulary as "green" at a ratio of  $\gamma$ . Subsequently, they elevate the likelihood of the tokens by boosting the logits of the softmax by  $\delta$ . Through this approach, at each token position, the probability of this matching between the seed and green tokens tends to increase.
+
+For a sentence with  $L$  tokens, it is viewed as a sample set of size  $L$ . Each token is one sample from the vocabulary. A non-watermarked sentence is expected to have  $\gamma L$  tokens showing this match, while the watermarked sentence is expected to have more. The watermark detection is approached as a  $z$ -test with null hypothesis that the text is non-watermarke. If the  $z$ -statistic is large, i.e. it is significantly different from the null hypothesis, the null hypothesis can be rejected and the text can be predicted as watermarked:
+
+$$
+z = \frac {(G - \gamma L)}{\sqrt {L \gamma (1 - \gamma)}}, \tag {1}
+$$
+
+where  $G$  is the number of tokens showing the matching between seed and the green list. Yoo et al. (2023) further expand this watermark of green and red list to more lists for multi-bit encoding.
+
+(Liu et al., 2023a) propose a semantic invariant method to watermark the generated text of LLM. However, their method employs two additional models, introducing redundant encoding processes in the text encoder, which can be time-consuming.
+
+# 3 Method
+
+In this section, we introduce the detailed design of SemaMark. We first present how to use the semantic information as the seed for watermark methods that are based on random partition of vocabulary in Section 3.1. Then in Section 3.2 and Section 3.3, we introduce the CL training scheme and the smoothed detection method for further improving the robustness, respectively.
+
+# 3.1 The framework of SemaMark
+
+As aforementioned, the existing watermark methods based on partition of vocabulary are susceptible to paraphrase. Paraphrase can easily change the previous tokens and disrupt the matching between tokens and the partition of vocabulary, without significantly affecting the semantic meaning. Thus, SemaMark uses the invariant semantics for watermarking by discretizing the embedding space to accommodate the minor perturbation of semantics and provide a stable mapping between semantics and vocabulary partition for the consequent token.
+
+However, discretization in a high-dimension space is intricate and non-trivial. Therefore, we first reduce the high-dimensional embedding space onto the 2D NE-Ring and then discretize via NE-Ring. The whole watermarking process is shown in Figure 1. SemaMark first reduces the dimension of the embedding space to obtain the discrete semantic values by two steps, i.e., weighted embedding pooling and discretizing by NE-Ring, and then uses the semantic value to partition the vocabulary. The logits of green list is shifted to increase the probability of matching between semantics and the consequent token for watermarking the LLM,  $f$ . In the following, we introduce more details about the two steps to obtain a stable semantic value.
+
+S1: weighted embedding pooling. To enhance the robustness, we aggregate the semantics of previous  $m$  tokens by the weighted mean pooling function  $P(\cdot)$  before dimension reduction, instead of using only one preceding token's embedding. In the ablation studies of Section 4.4, we show that the method has the best performance when  $m$  is neither too big nor too small. For the token sequence  $\{t_{i:i + m - 1}\}$  starting at position  $i$ , we use their semantics to generate the token in the  $m$  position,  $t_{i + m}$ . We denote their embeddings as  $\{e_{i:i + m - 1}\}$ .  $\{e_{i:i + m - 1}\}$  can be easily obtained from the LLM,  $f$ , that we want to watermark. Intuitively, in  $\{t_{i:i + m - 1}\}$ , the embeddings of tokens far from  $t_{i + m}$  contain semantic information that is more distant from  $t_{i + m}$  than the closer ones. The connection between distant tokens might be more possible to change after paraphrase compared with closer tokens. Thus, in the sequence  $\{t_{i:i + m - 1}\}$ , the embeddings of distant tokens might be less robust. To increase the robustness for the green list of the current token position  $t_{i + m}$  after paraphrase, the pooling embeddings should rely more on the closer tokens, therefore, we use a linear weight function to assign lower weights to tokens far from  $t_{i + m}$  and higher weights to those in closer proximity:
+
+$$
+P (\left\{\boldsymbol {e} _ {i + 1: i + m} \right\}) = \sum_ {j = 1} ^ {K} \frac {j + \frac {K}{2}}{w _ {\text {s u m}}} \boldsymbol {e} _ {i + j},
+$$
+
+where  $w_{\mathrm{sum}} = K^2 + K / 2$  is the sum of all weights. We denote the weighted output  $P(\{e_{i:i + m - 1}\}) \in \mathbb{R}^d$  as  $e_{P_{i,m}}$  for short. By pooling, more semantics are used for a seed, which enhances the robustness under paraphrase.
+
+S2: discretizing by NE-Ring. After aggregating the embeddings by weighted pooling, SemaMark uses MLP  $g_{\theta}$  to transform  $e_{P_{i,m}}$  to a normalized vector in 2D embedding space. The normalized vectors locate on a unit circle in the 2D space, which is named as Normalized Embedding Ring (NE-Ring). The discretization function,  $D(\cdot)$ , discretizes NE-Ring by equally segmenting into different sections. It takes the polar angle  $\phi$  of  $g_{\theta}(e_{P_{i,m}})$  as input and outputs the discretized semantic values  $a \in [K]$ , where  $[K] := \{1, 2, \dots, K\}$ .  $D(\cdot)$  is defined as
+
+$$
+D (\phi) = \left\lfloor \phi \frac {K}{2 \pi} \right\rfloor
+$$
+
+It first maps the input from  $[0,2\pi)$  to  $[0,K)$ , and then discretizes all the values in  $[i,i + 1)$  to  $i$ , for  $\forall i\in [K - 1]$ . Even though there could be subtle changes in semantics by paraphrase, the paraphrased  $\tilde{a}$  will likely locate in the discrete section  $[i,i + 1)$ . Some tokens may still have  $a\neq \tilde{a}$  if the normalized vector is close to the boundary of  $[i,i + 1)$ . Therefore, in Section 3.3, we introduce an offset detection to strengthen the tolerance for this mismatch and correct some unstable cases.
+
+With the two steps, we can get a stable discrete semantic value as the seed for the partition function to partition the vocabulary for the consequent token. Following Kirchenbauer et al. (2023a), the vocabulary is partitioned into green and red lists. We increase the logits of the tokens in the green list by  $\delta$  and recalculate the probability distribution based on the shifted logits. For each token to generate, we increase the possibility of the green list based on its previous  $m$  tokens' semantics. Thus, all the generated tokens will be likely to have this matching between the semantics and the consequent green token. By detecting the matching, we can discriminate whether a text is watermarked or not and then detect the LLM-generated contents effectively. Besides, SemaMark proposes two strategies to reduce the risk of being cracked by Contrastive Learning and further increase the robustness by the offset detection in the following sections.
+
+# 3.2 Training  $g_{\theta}$  by Contrastive Learning
+
+The MLP is expected to produce a uniform distribution of  $g_{\theta}(e_{P_{i,m}})$  on NE-Ring. If different semantics unevenly distributed on NE-Ring, the resulting discrete semantic values will be overly monotonous and the green list is more changeless. Consequently, the green list might be revealed by counting the token frequency, which compromises the concealment of watermark and leads to the risk of being cracked. Ideally, SemaMark should generate a wider variety of semantic values for different sentences, while each semantic value is robust and stable if its corresponding sentence is paraphrased. To achieve this goal, we propose to use Contrastive Learning to train MLP since Contrastive Learning has the property of uniformity that the data will be evenly distributed in the whole feature space (Wang and Isola, 2020). The uniform distribution can help the normalized vectors cover all the semantic values. As a result, NE-Ring can generate a wider variety of semantic values to prevent the watermark from being cracked.
+
+In Contrastive Learning, we first input the sentences into the model  $f$  to get a batch of sequences of  $m$  tokens and their pooling embeddings  $e_{P_{i,m}}$ , denoted as  $\{e_j\}$ , where  $j \in [B]$  and  $B$  is the batch size. To compose a contrastive loss, we construct the positive and negative pairs by a soft augmentation:
+
+$$
+\boldsymbol {e} _ {j + B} = \boldsymbol {e} _ {j} ^ {+} = \boldsymbol {e} _ {j} + \epsilon ,
+$$
+
+where  $\epsilon \sim \mathcal{N}(0, \sigma^2)$  is a Gaussian noise. The soft augmentation can simplify the construction of positive samples. With this soft augmentation, we can assign the samples sharing similar embeddings from the same sequence as positive pairs and samples from different sequences as negative pairs. This is consistent with our intuition that the paraphrased semantic embeddings will not change significantly and can remain robust. Then the contrastive loss is
+
+$$
+L _ {j} = - \log \frac {\exp \left(\sin \left(g _ {\theta} (\boldsymbol {e} _ {j}) , g _ {\theta} (\boldsymbol {e} _ {j} ^ {+})\right) / \tau\right)}{\sum_ {k \neq j , k \in [ 2 B ]} \exp \left(\sin \left(g _ {\theta} (\boldsymbol {e} _ {j}) , g _ {\theta} (\boldsymbol {e} _ {k})\right) / \tau\right)},
+$$
+
+where  $\mathrm{sim}(\cdot)$  is cosine similarity and  $\tau$  is the temperature. By Contrastive Learning, the output of reduced semantic embeddings can be evenly distributed in all of the space on NE-Ring, and cover all the discrete sections to improve the robustness of SemaMark.
+
+# 3.3  $Q$ -offset detection
+
+![](images/ab1779953e2c78717d5fb664ef364ddc96d3eaced36bd1c6c77bc584725ffac3.jpg)  
+Figure 2:  $Q$ -offset detection vs. existing detection
+
+Existing detection methods check the matching between partition seed and the consequent tokens in a one-to-one manner as shown in Figure 2(a). The detection method first recalculates the seed for each token position and gets the partition of the green
+
+list, and then checks whether the consequent token is in the partitioned green list token by token. In SemaMark, this strategy can be effective when the text is not paraphrased. However, after paraphrase, this detection could be suboptimal because the semantic values of some sequences which are close to the boundaries of the discrete section  $[i,i + 1)$  might change as shown in Figure 2(b). This is because the window of  $m$  tokens will slide token by token during the auto-regressive generation process, and the semantic change will also accumulate when the window is sliding. The semantic values closed to the boundary usually happen when the change accumulates to some extent. This change of boundary semantic values will lead to some mismatch and reduce the accuracy like  $\tilde{t}_5$  in Figure 2(b).
+
+To mitigate the influence of this error, we propose  $Q$ -offset detection. As shown in Figure 2(c), we offset the discrete seed by  $q$  tokens to detect the matching between semantics and the consequent tokens, where  $q \in \{-Q, -(Q - 1), \dots, 0, 1, \dots, Q\}$  and the sign of  $q$  indicates the direction of the offset. We choose the maximal  $z$ -statistic in different  $q$  as the  $Q$ -offset score. However,  $Q$ -offset detection will also increase the  $Q$ -offset score of non-watermark text, which indicates that the detected green word fraction  $\gamma$  of non-watermark text is higher. The  $\gamma$  in Eq. (1) is possibly inaccurate. Thus during generation, we set  $\gamma$  to a fixed value, while in detection process, we treat  $\gamma$  as a hyperparameter and use a validation set to determine its value in practice. In Section 4.4, we discuss the ablation studies of  $Q$ -offset and  $\gamma$  and show that  $Q$ -offset can impressively improve the detection performance with robustness.
+
+# 4 Experiment
+
+In this section, we conduct experiments to demonstrate the robustness of SemaMark. In Section 4.2, we demonstrate that its robustness is better than the baseline methods. In Section 4.3, we show that our watermark has almost no influence on the quality of generated texts. In Section 4.4, we use ablation studies to demonstrate the effectiveness of partition function and  $Q$ -offset detection, and show the sensitivity of the partition function. In Section 4.5 we visualize the distribution of NE-Ring and provide analysis on the feature distribution of Contractive Learning.
+
+<table><tr><td rowspan="2"></td><td rowspan="2">Paraphrase</td><td colspan="4">ROC-AUC</td><td colspan="4">F1 with best threshold</td></tr><tr><td>LeftHash</td><td>SelfHash</td><td>EXP-Edit</td><td>ours</td><td>LeftHash</td><td>SelfHash</td><td>EXP-Edit</td><td>ours</td></tr><tr><td rowspan="4">OPT-2.7B</td><td>No paraphrase</td><td>0.9913</td><td>0.9886</td><td>0.9799</td><td>0.9948</td><td>0.9921</td><td>0.9861</td><td>0.9708</td><td>0.9905</td></tr><tr><td>Translation</td><td>0.9091</td><td>0.8147</td><td>0.8749</td><td>0.9692</td><td>0.8456</td><td>0.7622</td><td>0.8157</td><td>0.9330</td></tr><tr><td>Dipper</td><td>0.9878</td><td>0.9728</td><td>0.9736</td><td>0.9911</td><td>0.9727</td><td>0.9400</td><td>0.9620</td><td>0.9701</td></tr><tr><td>GPT-3.5</td><td>0.9028</td><td>0.7908</td><td>0.9392</td><td>0.9406</td><td>0.8358</td><td>0.7378</td><td>0.8852</td><td>0.8902</td></tr><tr><td rowspan="4">OPT-6.7B</td><td>No paraphrase</td><td>0.9918</td><td>0.9930</td><td>0.9784</td><td>0.9949</td><td>0.9911</td><td>0.9863</td><td>0.9705</td><td>0.9858</td></tr><tr><td>Translation</td><td>0.8807</td><td>0.8098</td><td>0.8625</td><td>0.9308</td><td>0.8129</td><td>0.7468</td><td>0.8013</td><td>0.8882</td></tr><tr><td>Dipper</td><td>0.9904</td><td>0.9747</td><td>0.9728</td><td>0.9871</td><td>0.9786</td><td>0.9432</td><td>0.9620</td><td>0.9821</td></tr><tr><td>GPT-3.5</td><td>0.8990</td><td>0.7909</td><td>0.8996</td><td>0.9377</td><td>0.8300</td><td>0.7367</td><td>0.8354</td><td>0.8766</td></tr></table>
+
+Table 1: Watermark detection results under three paraphrases. (The best performance under paraphrase is bolded.)
+
+# 4.1 Experiment setups
+
+Backbone models and datasets. We test our method by watermarking two models, OPT-2.7B and OPT-6.7B (Zhang et al., 2022) which are referred to as the backbone models in following sections. For dataset, we use the news-like subset of C4 (Raffel et al., 2020), which covers a variety of topics. From the news-like subset of C4, we extract a training set, a validation set and a test set. For each sample, we use the first half of text as prompt to generate watermark sentences. More details can be found in Appendix A.
+
+Baseline methods. We compare our method with three baselines LeftHash, SelfHash (Kirchenbauer et al., 2023b) and EXP-Edit (Kuditipudi et al., 2023). LeftHash and SelfHash are two methods based on the partition of vocabulary using the hashes of tokens. EXP-Edit uses a private sequence to encode the watermark by changing the probability distribution of the sequence of tokens. More details on the implementation can be found in Appendix A.
+
+Paraphrase setups. We use three representative methods to paraphrase the watermarked text, round-trip translation (Tiedemann and Thottingal, 2020), Dipper (Krishna et al., 2023) and GPT-3.5. For round-trip translation, we first translate from English to another language and then transform back to English, such that some words and expressions will be changed because the translation is not an one-to-one mapping. For Dipper, we follow the parameter setting in Kirchenbauer et al. (2023b). For GPT-3.5, we use the prompt in Kirchenbauer et al. (2023b) to query GPT-3.5 for paraphrase. The examples of the three paraphrases can be found at Appendix B.
+
+Evaluation metrics and hyper-parameters. We use F1 score with best threshold and ROC-AUC to measure the performance of the watermark detection. All the metrics are calculated based on
+
+at least 500 watermarked samples and 500 non-watermark samples. The length of watermarked samples before paraphrase and non-watermark samples is  $200 \pm 25$ . In generation, we set  $\gamma = 1/4$  for LeftHash, SelfHash and SemaMark. In detection, we set  $\gamma = 1/3$  and  $\delta = 2$  based on the validation set in Section 4.4(b). In SemaMark, we set  $m = 15$ ,  $Q = 15$ ,  $K = 5$  for OPT-2.7B and  $K = 4$  for OPT-6.7B.
+
+# 4.2 Main Results
+
+In this subsection, we demonstrate the robustness of the proposed SemaMark under paraphrase by comparing it with three baseline methods on two backbone models. We first generate watermarked texts and use three paraphrase methods to remove the watermarks. The detection performance of both texts with and without paraphrase is reported in Table 1. As we can see, before paraphrase, all the watermarked methods have good detection performance. After paraphrase, SemaMark has the best detection performance most of the time across all the backbone models and all the paraphrase methods, which suggests that our method is more robust against paraphrase.
+
+In detail, by round-trip translation, the paraphrase reduces the detection ability of baseline methods effectively, while the watermark of SemaMark is robust. Under round-trip translation, the best ROC-AUC of baselines is 0.9091 on OPT-2.7B and 0.8807 on OPT-6.7B, respectively. But ROC-AUC of SemaMark is 0.9692 and 0.9308, which is at least 0.05 higher than all the baseline methods. Similarly, under paraphrase of GPT-3.5, SemaMark is better than all the baselines. The best baseline performance under GPT-3.5 is 0.9392 in ROC-AUC on OPT-2.7B and 0.8990 in ROC-AUC on opt-6.7B, but SemaMark has higher AUC-ROC of 0.9406 and 0.9377. For Dipper, we note that all methods are robust to Dipper since it does not sig
+
+nificantly reduce the detection performance. However, SemaMark is still one of the most robust. On OPT-2.7B, it performs best in ROC-AUC, while on OPT-6.7B, it has the best F1 score. From Table 1, the results show an obvious improvement of SemaMark in robustness. This implies that using semantics as the seed for the partition function is effective under paraphrase.
+
+# 4.3 Text Quality
+
+![](images/30249de8e3a17d6dd17b708f38b4bd694f4ebc5d9e4d1b19b32f8da2539e16f5.jpg)  
+(a) OPT-2.7B
+
+![](images/7484c078ce9cb79bd8c654e25b90293de7e490bee8bb54df8866bd5b8fbeb769.jpg)  
+(b) OPT-6.7B  
+Figure 3: Text quality (perplexity)
+
+Watermark should not compromise the generation quality of LLMs. In this subsection, we compare the text quality by calculating perplexity and demonstrate that our watermark has almost no influence on the generated quality. Perplexity measures the likelihood that a sentence is generated by one model. Lower perplexity means the watermarked text is more predictable. In other words, it is more consistent with the reasoning of the given model. In Figure 3, we use OPT-6.7B with no watermark to get perplexity for all the watermarked methods. All the results in Figure 3 are calculated without paraphrase, because the generation quality of text is not related to paraphrase. From Figure 3a on OPT-2.7B, we can see that our watermark, LeftHash and SelfHash have almost no influence on the generation quality. They has perplexity at around 6 which is similar as the generated text without watermark. Instead, EXP-Edit has much higher perplexity, which means that EXP-Edit changes the generated text in an aggressive way and much reduces the generation quality after watermarking. This is probably because EXP-Edit adjusts the logits on the whole vocabulary. From Figure 3b, we can draw similar conclusions for OPT-6.7B. EXP-Exit also increases the perplexity by around 10, while the average perplexity of LeftHash, SelfHash and ours is around 1 higher than the non-watermarked generated text. In summary, our SemaMark can
+
+![](images/f1fedbbb0715eac672aea6cc7255b03e27a0366f52cf5acbf89e20aac0aa7454.jpg)  
+Figure 4: ROC-AUC and  $m$
+
+![](images/13a767fdb4adaf1727203221315351f0b4c8ead709333794399687afdb5a684b.jpg)  
+(a) ROC-AUC and offset  $Q$
+
+![](images/1af1c8e4ea05979cb9a711b77fbd619ea31a156df58904c4d9a5119557fb0e15.jpg)  
+(b) ROC-AUC and  $\gamma$  
+Figure 5: Text quality (perplexity)
+
+keep the quality and robustness simultaneously.
+
+# 4.4 Ablation Study
+
+In this subsection, we study the influence of the length of the sequence we use for generating one semantic value and the sensitivity of the partition function.
+
+a) Length of previous sequence tokens,  $m$ . In the first step of SemaMark, i.e., weighted embedding pooling, we use the semantic of the previous  $m$  tokens to get the more stable embedding. But if the length of the sequence is too long, it will also hurt the robustness. In Figure 4, we test watermark on OPT-2.7B with different  $m$  and draw the ROC-AUC. The results show that before  $m = 15$ , ROC-AUC is in the trend of increase as the  $m$  changes. But when  $m > 15$ , ROC-AUC becomes fluctuating. It is possibly because that the distant tokens will include more change after paraphrase as we mentioned in Section 3.1. Another possible reason is that in the beginning of generation for the first  $m$  tokens, the number of previous tokens is smaller than  $m$  and NE-Ring can only use the embeddings of limited tokens for prediction, which may be unstable. Thus, too long or too short sequence will hurt the robustness of SemaMark against paraphrase. In our experiments, we choose  $m = 15$  for all the settings.
+
+b)  $Q$ -offset detection In this subsection, we show that the effectiveness of the proposed  $Q$ -offset detection. In Figure 5a, we demonstrate the change
+
+of ROC-AUC of SemaMark with different  $Q$  in offset detection under three different paraphrases.  $Q$ -offset detection searches the highest  $z$ -statistics from  $-Q$  to  $Q$  as the  $Q$ -offset score. From Figure 5a, we can see that when  $Q$  increases, ROC-AUC first increases and decreases after  $Q$  is around 15. When  $Q < 15$ , the offset can help correct the errors of semantic values close to the boundary. Compared with detection without offset, i.e.  $Q = 0$ , ROC-AUC of SemaMark is much better, which means that the offset can help to solve the errors of semantic values around the boundaries that are more vulnerable to paraphrases. When  $Q > 15$ , the correction of this error is limited, because the offset will also increase the  $Q$ -offset score of negative samples as it also searches the highest  $z$ -statistics of negative samples. On the other hand, the computation cost will also increase if  $Q$  is too large because it has to search more possible  $q$ . In practice, we set  $Q = 15$  in all the experiments, which can effectively reduce the influence of the errors of semantic values at the boundaries.
+
+Since the  $Q$ -offset detection searches the highest green word fraction, the fraction of green list word of non-watermarked text will be higher than the  $\gamma$  that we used to randomly select the green list. Thus, it is not accurate to use the original  $\gamma$  for  $z$ -statistics. We treat  $\gamma$  as a hyper-parameter and use a validation set to select its value. As shown in Figure 5b, the detection performance of SemaMark under paraphrases of Dipper and GPT-3.5 will reach the highest when  $\gamma$  is around  $1/3$ , while it will continue to increase under paraphrase. In practice, we set  $\gamma = 1/3$  for  $Q$ -offset detection.
+
+c) Sensitivity of partition function. As we mentioned, the partition function is sensitive to any change of the input because it only uses the input as the seed of the random generator. To validate its sensitivity to continuous embeddings, we adopt the embedding vector as the input to show that, with tiny change of the embeddings, the partition of vocabulary can be very different. We propose a hash method based on md5sum (Deepakumara et al., 2001) to adopt the partition function by transforming the continuous embeddings to an integral seed. We use 1000 sequences to test the sensitivity. For each sequence embedding, we first get a green list from the partition function. Then we change one dimension of the embedding by only 1e-5 to get a new partition result. The overlapping of the green list before and after changing is  $24.99\%$  on the av
+
+<table><tr><td rowspan="2"></td><td colspan="3">ROC-AUC</td><td colspan="3">F1 with best threshold</td></tr><tr><td>LeftHash</td><td>SelfHash</td><td>ours</td><td>LeftHash</td><td>SelfHash</td><td>ours</td></tr><tr><td>LLaMA-7B</td><td>0.819</td><td>0.838</td><td>0.846</td><td>0.748</td><td>0.774</td><td>0.781</td></tr><tr><td>LLaMA2-7B</td><td>0.811</td><td>0.841</td><td>0.872</td><td>0.749</td><td>0.773</td><td>0.810</td></tr></table>
+
+Table 2: Watermark detection results under different model size.)
+
+erage of 1000 sequences. It is consistent with  $\gamma$  we use to watermark, because the random partition with the changed embedding is independent from the original one. It means the partition function is sensitive to any small change in its input. Instead, after we use NE-Ring to discretize the embeddings, the overlapping of green list after changing embeddings by 1e-5 is  $100\%$ , which means the discretization can effectively handle this change. In practice, SemaMark can provide the tolerance that is much larger than 1e-5, which makes the watermark more robust under paraphrase. With the improvement of  $Q$ -offset, the detection of SemaMark is more robust and effective.
+
+$d$  ) Model size. To show the robustness of our method on different model sizes, in this section, we also test the watermark under round-trip translation paraphrase on LLaMA-7B and LLaMA2-7B, which have larger size and different architectures. As indicated in Table 3, our approach consistently exhibits the highest robustness against paraphrasing. Specifically, in the LLaMA2-7B model, SemaMark significantly outperforms the baseline models, achieving an increase of 0.06 and 0.03 in ROC-AUC. Similarly, in the LLaMA-7B model, our method shows superior performance with an increase of 0.027 and 0.009 in ROC-AUC.
+
+# 4.5 Distribution on NE-Ring based on CL
+
+![](images/ab40d2ca72fb44afadf6354b472756c2c23770329f5cd7f68c9e34bf85f09458.jpg)  
+(a) NE-Ring  
+Figure 6: Visualization of NE-Ring
+
+![](images/c7d99a18669bf41283712fe7dd29af4c993660d36aa002527546083eba306c3a.jpg)  
+(b) Distribution on  $\phi$
+
+In this subsection, we demonstrate that Contrastive Learning can help evenly distribute the semantics on the NE-Ring. The even distribution can help the sequences reach all possible semantic values and provide more diverse semantic values
+
+to prevent the watermark from being cracked by counting token frequency. In Figure 6a, we use Gaussian density estimation (Chen, 2017) to get the distribution of the semantics on the NE-Ring before discretization. We use different colors to show the density. The NE-Ring in Figure 6a shows that, the distribution is uniform. All the density is between 0.052 and 0.054. We further plot the density based on the polar angle  $\phi$  in Figure 6b where the density has almost no change on all the polar angle from 0 to  $2\pi$ . This implies that the training based on Contrastive Learning can ensure the semantics will reach all possible discrete values. It can prevent the case where the discrete values will gather in some discrete sections and produce monotonous vocabulary partitions. As a result, it can protect the watermark from being cracked by counting token frequency.
+
+# 5 Conclusion
+
+In this paper, we use the semantic information for watermarking to enhance the robustness against paraphrase. The existing watermark methods use the matching between the previous tokens and the partition vocabulary. This matching can be easily broken by paraphrase. However, we construct the mapping between the semantics and the vocabulary. In this way, the semantics will stay stable under paraphrase and the robustness of watermark can be enhanced. To make use of semantics, we propose SemaMark to discretize the embedding space on NE-Ring and propose a training method based on CL. In addition, we use  $Q$ -offset detection to further advance the robustness by increasing the tolerance of the semantic values close to the discrete boundary. In experiments, we demonstrate our method can perform much better compared with baseline methods under paraphrase while having little influence on the generation quality.
+
+# 6 Limitations
+
+In some cases, the customers may rely on some API-based LLMs and do not have the access to the embeddings and the permission to modify the logits during generation. Although our watermark method can effectively detect the LLM-generated content and increase the detection success rate under paraphrase, it is not applicable for black-box LLMs. The second weakness of our method is that the NE-Ring is dependent on the semantic embedding of LLMs. For each LLM, we need to train a
+
+specialized EN-Ring, which might be inflexible if we want to produce a general model for NE-Ring or fine-tune the LLMs. Despite the weaknesses, our method is successful in the problem of robustness under paraphrase. In the future work, we will continue to extent our method into black-box LLMs and a universal model that does not require customized training for various specific LLMs.
+
+Potential risk. Our discussion about the robustness might provide motivation for the attackers to find other methods like adaptive attack. Although we provide robustness under paraphrase, if the unauthorized people propose possible attack method focusing on the green-list based watermark from other perspectives, the detection rate for LLM-generated texts are still possible to be reduced.
+
+# 7 Acknowledgements
+
+Jie Ren, Han Xu, Yingqian Cui, and Jiliang Tang are supported by the National Science Foundation (NSF) under grant numbers CNS 2246050, IIS1845081, IIS2212032, IIS2212144, IOS2107215, DUE 2234015, DRL 2025244 and IOS2035472, the Army Research Office (ARO) under grant number W911NF-21-1-0198, the Home Depot, Cisco Systems Inc, Amazon Faculty Award, Johnson&Johnson, JP Morgan Faculty Award and SNAP.
+
+# References
+
+Steven Bird, Ewan Klein, and Edward Loper. 2009. Natural language processing with Python: analyzing text with the natural language toolkit. "O'Reilly Media, Inc."  
+Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR.  
+Yen-Chi Chen. 2017. A tutorial on kernel density estimation and recent advances. Biostatistics & Epidemiology, 1(1):161-187.  
+Antonia Creswell, Murray Shanahan, and Irina Higgins. 2022. Selection-inference: Exploiting large language models for interpretable logical reasoning. arXiv preprint arXiv:2205.09712.  
+Janaka Deepakumara, Howard M Heys, and R Venkatesan. 2001. FPGA implementation of md5 hash algorithm. In Canadian Conference on Electrical and Computer Engineering 2001. Conference Proceedings (Cat. No. 01TH8555), volume 2, pages 919-924. IEEE.
+
+Tianyu Gao, Xingcheng Yao, and Danqi Chen. 2021. Simcse: Simple contrastive learning of sentence embeddings. In 2021 Conference on Empirical Methods in Natural Language Processing, EMNLP 2021, pages 6894-6910. Association for Computational Linguistics (ACL).  
+John Giorgi, Osvald Nitski, Bo Wang, and Gary Bader. 2021. Declutr: Deep contrastive learning for unsupervised textual representations. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 879-895.  
+Biyang Guo, Xin Zhang, Ziyuan Wang, Minqi Jiang, Jinran Nie, Yuxuan Ding, Jianwei Yue, and Yupeng Wu. 2023. How close is chatgpt to human experts? comparison corpus, evaluation, and detection. arXiv preprint arXiv:2301.07597.  
+Charles R. Harris, K. Jarrod Millman, Stefan J. van der Walt, Ralf Gommers, Pauli Virtanen, David Cournaepau, Eric Wieser, Julian Taylor, Sebastian Berg, Nathaniel J. Smith, Robert Kern, Matti Picus, Stephan Hoyer, Marten H. van Kerkwijk, Matthew Brett, Allan Haldane, Jaime Fernandez del Rio, Mark Wiebe, Pearu Peterson, Pierre Gerard-Marchant, Kevin Sheppard, Tyler Reddy, Warren Weckesser, Hameer Abbasi, Christoph Gohlke, and Travis E. Oliphant. 2020. Array programming with NumPy. Nature, 585(7825):357-362.  
+Jean Kaddour, Joshua Harris, Maximilian Mozes, Herb Bradley, Roberta Raileanu, and Robert McHardy. 2023. Challenges and applications of large language models. arXiv preprint arXiv:2307.10169.  
+John Kirchenbauer, Jonas Geiping, Yuxin Wen, Jonathan Katz, Ian Miers, and Tom Goldstein. 2023a. A watermark for large language models. arXiv preprint arXiv:2301.10226.  
+John Kirchenbauer, Jonas Geiping, Yuxin Wen, Manli Shu, Khalid Saifullah, Kezhi Kong, Kasun Fernando, Aniruddha Saha, Micah Goldblum, and Tom Goldstein. 2023b. On the reliability of watermarks for large language models. arXiv preprint arXiv:2306.04634.  
+Sarah Kreps, R Miles McCain, and Miles Brundage. 2022. All the news that's fit to fabricate: A-generated text as a tool of media misinformation. Journal of experimental political science, 9(1):104-117.  
+Kalpesh Krishna, Yixiao Song, Marzena Karpinska, John Wieting, and Mohit Iyyer. 2023. Paraphrasing evades detectors of ai-generated text, but retrieval is an effective defense. arXiv preprint arXiv:2303.13408.  
+Rohith Kuditipudi, John Thickstun, Tatsunori Hashimoto, and Percy Liang. 2023. Robust distortion-free watermarks for language models. arXiv preprint arXiv:2307.15593.
+
+Bohan Li, Hao Zhou, Junxian He, Mingxuan Wang, Yiming Yang, and Lei Li. 2020. On the sentence embeddings from pre-trained language models. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 9119-9130.  
+Linyang Li, Pengyu Wang, Ke Ren, Tianxiang Sun, and Xipeng Qiu. 2023. Origin tracing and detecting of llms. arXiv preprint arXiv:2304.14072.  
+Aiwei Liu, Leyi Pan, Xuming Hu, Shiao Meng, and Lijie Wen. 2023a. A semantic invariant robust watermark for large language models. arXiv preprint arXiv:2310.06356.  
+Aiwei Liu, Leyi Pan, Yijian Lu, Jingjing Li, Xuming Hu, Lijie Wen, Irwin King, and Philip S Yu. 2023b. A survey of text watermarking in the era of large language models. arXiv preprint arXiv:2312.07913.  
+Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. 2022. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521.  
+Eric Mitchell, Yoonho Lee, Alexander Khazatsky, Christopher D Manning, and Chelsea Finn. 2023. Detectgpt: Zero-shot machine-generated text detection using probability curvature. arXiv preprint arXiv:2301.11305.  
+Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67.  
+Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3982-3992.  
+Jie Ren, Han Xu, Pengfei He, Yingqian Cui, Shenglai Zeng, Jiankun Zhang, Hongzhi Wen, Jiayuan Ding, Hui Liu, Yi Chang, et al. 2024. Copyright protection in generative ai: A technical perspective. arXiv preprint arXiv:2402.02333.  
+Vinu Sankar Sadasivan, Aounon Kumar, Sriram Balasubramanian, Wenxiao Wang, and Soheil Feizi. 2023. Can ai-generated text be reliably detected? arXiv preprint arXiv:2303.11156.  
+Joni Salminen, Chandrashekhar Kandpal, Ahmed Mohamed Kamel, Soon-gyo Jung, and Bernard J Jansen. 2022. Creating and detecting fake reviews of online products. Journal of Retailing and Consumer Services, 64:102771.
+
+Ruixiang Tang, Yu-Neng Chuang, and Xia Hu. 2023. The science of detecting llm-generated texts. arXiv preprint arXiv:2303.07205.  
+Jörg Tiedemann and Santhosh Thottingal. 2020. OPUSMT — Building open translation services for the World. In Proceedings of the 22nd Annual Conference of the European Association for Machine Translation (EAMT), Lisbon, Portugal.  
+Saranya Venkatraman, Adaku Uchendu, and Dongwon Lee. 2023. Gpt-who: An information density-based machine-generated text detector.  
+Pengyu Wang, Linyang Li, Ke Ren, Botian Jiang, Dong Zhang, and Xipeng Qiu. 2023. Seqxgpt: Sentence-level ai-generated text detection.  
+Tongzhou Wang and Phillip Isola. 2020. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International Conference on Machine Learning, pages 9929-9939. PMLR.  
+Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35:24824-24837.  
+Junchao Wu, Shu Yang, Runzhe Zhan, Yulin Yuan, Derek F Wong, and Lidia S Chao. 2023. A survey on llm-gernerated text detection: Necessity, methods, and future directions. arXiv preprint arXiv:2310.14724.  
+Frank F Xu, Uri Alon, Graham Neubig, and Vincent Josua Hellendoorn. 2022. A systematic evaluation of large language models of code. In Proceedings of the 6th ACM SIGPLAN International Symposium on Machine Programming, pages 1-10.  
+Han Xu, Jie Ren, Pengfei He, Shenglai Zeng, Yingqian Cui, Amy Liu, Hui Liu, and Jiliang Tang. 2023. On the generalization of training-based chatgpt detection methods. arXiv preprint arXiv:2310.01307.  
+KiYoon Yoo, Wonhyuk Ahn, Jiho Jang, and Nojun Kwak. 2023. Robust multi-bit natural language watermarking through invariant features. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2092-2115.  
+Rowan Zellers, Ari Holtzman, Hannah Rashkin, Yonatan Bisk, Ali Farhadi, Franziska Roesner, and Yejin Choi. 2019. Defending against neural fake news. Advances in neural information processing systems, 32.  
+Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. 2022. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068.
+
+Xuandong Zhao, Prabhanjan Ananth, Lei Li, and Yu-Xiang Wang. 2023. Provable robust watermarking for ai-generated text. arXiv preprint arXiv:2306.17439.
+
+# A More details on experimental settings
+
+All the baseline models, backbone models and datasets we use are open source and available for academic purpose. For backbone models, we use the open-sourced model from Huggingface<sup>1</sup>. The implementation is based on Pytorch<sup>2</sup> framework and also depend on packages including NLTK (Bird et al., 2009) and Numpy (Harris et al., 2020). For baseline methods, we use the released official code from the authors. For paraphrase models, we use OPUS-MT translation model and Dipper on Huggingface repository<sup>3</sup>, and API of ChatGPT<sup>4</sup>.
+
+# B Examples of paraphrases
+
+In Table 3, we provide the examples of three paraphrases.
+
+<table><tr><td></td><td>Original</td><td>Paraphrased</td></tr><tr><td>Translation</td><td>The course ran from Feb. 16 to Feb. 18 and was designed to address officers&#x27; response to critical incidents and to enhance officer safety. &quot;I am excited by the training and professionalism demonstrated by the officers at the training session. They were receptive and focused on learning new techniques and approaches to policing.&quot; Com-missioner Walwyn added.</td><td>The training, which was held from 16 February to 18 February, was aimed at discussing police responses to major incidents and enhancing police security. &quot; I am excited about the training and professionalism shown by police officers in the course, who are willing to accept and focus on new policing techniques and methods. &quot; Com-missioner Walwyn adds.</td></tr><tr><td>GPT</td><td>She tells him she wants to protect him. She asks what else she can do. Stefan tells her not to trust Chad. She vows she won&#x27;t. Stefan offers to get rid of Chad for her if she wants. She says she wants Chad and his family gone. She wants to have his babies. She tells him to get out of her way. She doesn&#x27;t know if she can love him, but she wants his babies. She wants to be a mother. Stefan tells her she deserves a real man. He&#x27;ll love her and protect her. She says she wants him and his babies. She wants to have his babies. Stefan says she deserves a real man. He&#x27;ll love her and protect her. She says she wants him and his babies. She wants to have his babies. Stefan says she deserves a real man. He&#x27;ll love her and protect her. She says she wants him and his babies. She wants to have his babies. Stefan says she deserves a real man. He&#x27;ll love her and protect her. She says she wants him and his babies.</td><td>According to Reuters, China has already taken steps to solidify its presence in the South China Sea. It has constructed an airstrip on one island and In a tense conversation, a woman expresses her desire to protect a man, Stefan, and seeks his advice on how to do so. Stefan advises her not to trust a person named Chad, to which she firmly pledges her distrust. Stefan offers to eliminate Chad from her life if she desires, and she responds by expressing her wish for Chad and his family to disappear entirely. In a surprising revelation, she declares her desire to have Stefan&#x27;s children, emphasizing her longing to become a mother. Despite uncertain feelings of love towards Stefan, she insists on wanting his babies and urges him to step aside. Stefan, in turn, assures her that she deserves a genuine partner who will love and safeguard her. Yet, undeterred, she asserts her desire for both Stefan and his offspring, reiterat-ing her ambition to have his babies.</td></tr><tr><td>DIPPER</td><td>Kewanee got the ball again with 7.2 seconds to go, but couldn&#x27;t get a shot off. &quot;I think it was something we needed to have a plan for and it was one that worked in the end. We did it well enough,&quot; Keene said. &quot;We were trying to get to the rim. A lot of times when you are in a situation like that and you can&#x27;t score, it&#x27;s better to just foul.&quot; Macomb was 12-for-27 from behind the arc (42 percent). The Grinnell offense is centered around perimeter shooters. The two perimeter players for Macomb are Emerick and Reilly Rieder. The next closest shooter on the team is Rieder with 10 points. Nolan led Kewanee with 15 points. Macomb, now 14-5, now travels to Jacksonville on Thursday to play Jacksonville</td><td>Kewanee had the ball again with 7.2 seconds left but couldn&#x27;t get off a shot. &quot;I think that&#x27;s something we needed to have a plan for and in the end, it worked,&quot; coach Lambert said. &quot;We did it well enough,&quot; Keene said. &quot;We were trying to get to the rim. Often when you are in a situation like that and you can&#x27;t score, it&#x27;s better to foul.&quot; Macomb was 12 for 27 from beyond the arc (42 percent). The Grinnell offense is based on sharpshooting players. Macomb&#x27;s two shooters are Emerick and Rieder. Rieder has ten points. Nolan led Kewanee with 15 points. Macomb, now 14-5, will play at Jacksonville Thursday.</td></tr></table>
+
+Table 3: Paraphrase examples.

2024/A Robust Semantics-based Watermark for Large Language Model against Paraphrasing/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f68f08d5aeac4e400ce37db40ec89b08bff49db947ad93d863c6802000de70b7
+size 727507

2024/A Study on Scaling Up Multilingual News Framing Analysis/a4b2e349-9396-4656-a45a-c4d7d00460a3_content_list.json

@@ -0,0 +1,1768 @@
+[
+    {
+        "type": "text",
+        "text": "A Study on Scaling Up Multilingual News Framing Analysis",
+        "text_level": 1,
+        "bbox": [
+            186,
+            90,
+            811,
+            112
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Syeda Sabrina Akter and Antonios Anastasopoulos  \nDepartment of Computer Science, George Mason University  \n{sakter6, antonis}@gmu.edu",
+        "bbox": [
+            253,
+            145,
+            747,
+            195
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            260,
+            341,
+            275
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Media framing is the study of strategically selecting and presenting specific aspects of political issues to shape public opinion. Despite its relevance to almost all societies around the world, research has been limited due to the lack of available datasets and other resources. This study explores the possibility of dataset creation through crowdsourcing, utilizing non-expert annotators to develop training corpora. We first extend framing analysis beyond English news to a multilingual context (12 typologically diverse languages) through automatic translation. We also present a novel benchmark in Bengali and Portuguese on the immigration and same-sex marriage domains. Additionally, we show that a system trained on our crowdsourced dataset, combined with other existing ones, leads to a 5.32 percentage point increase from the baseline, showing that crowdsourcing is a viable option. Last, we study the performance of large language models (LLMs) for this task, finding that task-specific fine-tuning is a better approach than employing bigger non-specialized models. $^{1}$",
+        "bbox": [
+            141,
+            292,
+            463,
+            634
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            650,
+            260,
+            665
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "News framing refers to the power of the news media to define and interpret events, issues, and policies by emphasizing certain aspects while downplaying or excluding others. According to Entman (1993), it can \"make a piece of information more noticeable, meaningful, or memorable to audiences\". It plays a crucial role in influencing how people interpret and react to information presented in news articles. The language used in news media can shape public opinion and reveal biases and agendas, which can ultimately shape the way people understand and react to current events.",
+        "bbox": [
+            112,
+            678,
+            490,
+            870
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "- A presidente do País, Tsai Ing-wen, également se pronunciou a favor da lei e pediu acos legisladores empenho com a pauta Political. \"Temos a opportunidade de fazer-history e做不到 ao mundo que valeores progressistas podemcriar raízes nas sociedades da Ásia Oriental\", affirmou Fairness and Equality.",
+        "bbox": [
+            517,
+            268,
+            885,
+            355
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Figure 1: The image illustrates the process of framing in Portuguese at the sentence level, showcasing how specific language for each sentence strategically shape a Political and Equality narrative in the same article.",
+        "bbox": [
+            507,
+            372,
+            882,
+            431
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Traditionally, framing analysis has relied on manual annotation by linguists, social studies experts, and trained annotators, lacking the potential of AI-driven systems leading to a rather limited explorations of automating framing analysis. Moreover, existing studies have been restricted primarily to English-only data, leaving a gap in research concerning multilingual and low-resource contexts.",
+        "bbox": [
+            507,
+            437,
+            885,
+            565
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Our work focuses on employing NLP techniques for the framing analysis task to automate the analysis process, extract insights from large datasets efficiently, and identify patterns in the language used in news media. To address these challenges, Boydstun et al. (2014) introduced a codebook, Policy Frames Codebook, based on which the Media Frames Corpus (MFC; Card et al., 2015) was created. This dataset is comprised broad categories of common policy frames and annotations of US news articles. However, the availability of such datasets in languages beyond English remains limited.",
+        "bbox": [
+            507,
+            565,
+            885,
+            759
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Getting a higher volume of higher quality data (such as, MFC) is time and resource intensive. Hence, we study the alternative of gathering a high volume of comparatively lower quality but easy-to-collect data. We achieve this through crowdsourcing and automatic translation techniques. We also examine the combination of lower and higher quality data.",
+        "bbox": [
+            507,
+            760,
+            885,
+            888
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In this study, we first introduce a new crowdsourced dataset: Student-sourced Noisy Frames",
+        "bbox": [
+            507,
+            889,
+            884,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{1}$ Code and Dataset available here: https://github.com/syedasabrina/ Scaling-up-multilingual-framing-analysis.git",
+        "bbox": [
+            112,
+            882,
+            487,
+            920
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "4156",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: NAACL 2024, pages 4156-4173 June 16-21, 2024 ©2024 Association for Computational Linguistics",
+        "bbox": [
+            218,
+            945,
+            778,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Corpus (SNFC). We have achieved time and cost efficiency by involving a large number of semitrained annotators for the data collection and annotation process of the corpus. SNFC covers immigration and same-sex marriage domains and includes novel benchmark test sets in Bengali and Portuguese, offering new perspectives in these languages. Additionally, we automatically expand multilinguality to the task by translating the MFC and SNFC to 12 more languages. We show that a neural classifier trained on the combination of both MFC and SNFC yields significant performance improvements, both in English as well as in a multilingual setting. Finally, we explore generative large language models, such as LLaMA (Touvron et al., 2023), to study their efficacy for this task.",
+        "bbox": [
+            115,
+            84,
+            485,
+            340
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our findings show that neural models trained on SNFC can reach the performance levels of those trained on high quality data (i.e., MFC). Going further, we find that the combination of expert and non-expert annotated data (i.e. MaSNFC+MFC) outperforms just MFC, which provides a path towards expanding coverage without the need for expensive expert annotations.",
+        "bbox": [
+            115,
+            343,
+            485,
+            469
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            115,
+            488,
+            267,
+            502
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Framing analysis provides valuable insights into different perspectives on news topics across various countries and languages. However, there is a notable lack of research and annotated corpora for framing analysis in languages other than English. This limitation hinders our understanding of media framing in different parts of the world and other societies' opinion regarding specific issues. To address this gap, a multilingual approach is essential in analyzing media framing across diverse linguistic and cultural contexts. Ali and Hassan (2022) provide a comprehensive survey of the framing analysis task, focusing specifically on studies in English datasets exploring various approaches and techniques employed in framing analysis.",
+        "bbox": [
+            115,
+            517,
+            485,
+            758
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Two prominent datasets used for framing analysis are the Media Frames Corpus (MFC; Card et al., 2015) and the Gun Violence Frames Corpus (GVFC; Liu et al., 2019). The MFC, annotated according to the guidelines provided in the codebook of Boydstun et al. (2014), covers 6 different political issues including immigration, same-sex marriage, and gun violence, among others. It includes both article headlines and news texts, providing a broader and more comprehensive dataset.",
+        "bbox": [
+            115,
+            760,
+            485,
+            919
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "On the other hand, the GVFC focuses solely on the topic of gun violence, with 10 manually annotated frames defined in a different codebook, and it only includes article headlines.",
+        "bbox": [
+            512,
+            84,
+            880,
+            146
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Akyurek et al. (2020) extended the GVFC by curating headlines in German, Turkish, and Arabic following the same process as the original dataset from the respective news websites, specifically targeting keywords related to gun violence and mass shootings. The frames used in the multilingual datasets remained consistent with those in the GVFC, and is the one of the few multilingual sources for this task. Additionally, the Australian Parliamentary Speeches (APS) dataset (Khanehzar et al., 2019) offers another perspective on framing analysis, as it consists of transcripts speeches related to same-sex marriage bills presented in the Australian Parliament. Although the APS dataset focuses on data from a country other than the United States, it is still limited to English language texts, which narrows the scope of the framing analysis task.",
+        "bbox": [
+            512,
+            148,
+            882,
+            436
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The MFC has served as a valuable resource in various framing-related studies. For example, it was used to develop a semi-supervised model by extracting a Russian lexicon from their Russian test corpora which consists of news articles sourced from reputable Russian newspapers (Field et al., 2018). In a different vein, Naderi and Hirst (2017) used it to benchmark sentence-level classification tasks, employing LSTM, BiLSTM, and GRU-based systems. Considering the significant contributions of this corpus to the field, we have incorporated it into our system for training and evaluation purposes, alongside our SNFC dataset.",
+        "bbox": [
+            512,
+            438,
+            882,
+            646
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Several studies have employed various techniques such as topic modeling (DiMaggio et al., 2013; Roberts et al., 2014; Nguyen, 2015), cluster analysis (Burscher et al., 2016), and neural networks (Naderi and Hirst, 2017; Khanehzar et al., 2019; Mendelsohn et al., 2021; Kwak et al., 2020) to construct systems for framing analysis. These investigations have consistently demonstrated that leveraging state-of-the-art pre-trained models based on transformers (Devlin et al., 2019; Zhuang et al., 2021; Conneau et al., 2020) is a highly effective approach, yielding significantly improved results compared to other techniques. In our study, we follow the state of the art and build models similar to those employed by Liu et al. (2019) and Khanehzar et al. (2019).",
+        "bbox": [
+            512,
+            648,
+            882,
+            903
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We also investigated crowdsourcing methods",
+        "bbox": [
+            531,
+            904,
+            880,
+            919
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "4157",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "which, as defined by Howe (2006), is an online, distributed problem-solving and production model that leverages the collective intelligence of online communities for specific goals. This technique aims to tap into the global talent pool, accelerating innovation and problem-solving across various domains. Hossain and Kauranen (2015) provide a comprehensive literature review, identifying numerous crowdsourcing methods, which emphasizes the difficulty of generalizing these methods due to their diversity and application-specific nature. However, the widespread use of these methods demonstrates versatility and adaptability of different crowdsourcing methods. Zhao and Zhu (2014) suggest that future research should focus on standardizing crowdsourcing processes to enhance efficiency and effectiveness. This indicates an increasing realization of the necessity to codify crowdsourcing approaches, notwithstanding their inherent variability.",
+        "bbox": [
+            110,
+            84,
+            492,
+            391
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Dataset Creation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            404,
+            294,
+            419
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this section, we present our methodology for curating SNFC training dataset through crowdsourcing (§3) and outline the process of extending the dataset to incorporate multilinguality (§3). Lastly, we introduce our innovative Portuguese and Bengali benchmarks, highlighting their significance in the context of this study (§3).",
+        "bbox": [
+            112,
+            432,
+            489,
+            544
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "SNFC Training Corpus To construct the crowdsourced training portion of the SNFC, we turned to students at George Mason University. In particular, this was done as part of an in-class assignment for a graduate-level natural language processing class with about 80 students involved.[2]",
+        "bbox": [
+            112,
+            556,
+            489,
+            651
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The students were presented with the challenge of building a Media Frames Analysis system (effectively, a sentence-level neural classifier), without having access to significant amounts of data. In particular, the students were provided only with a description of the codebook of Boydstun et al. (2014) presented in Table 5, along with 250 sentence-level examples called the seed dataset from the MFC corpus sampled so that all 15 frame dimensions were present.",
+        "bbox": [
+            112,
+            653,
+            489,
+            813
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The codebook and the samples were meant to facilitate the annotators' understanding of the task. The only other information available to them was that their final systems would be evaluated on multiple languages (see §3) on the immigration and",
+        "bbox": [
+            112,
+            815,
+            489,
+            896
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "same-sex marriage domains.3",
+        "bbox": [
+            509,
+            84,
+            731,
+            99
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The students were first tasked with procuring 150 new sentences each, from any source and in any language, and label them, according to the codebook, to be used as their \"first\" training set. They then had to produce an additional 150 sentences which would then be annotated by two of their peers (so that we will be able to measure inter-annotator agreement). Any label disagreements were resolved by the students, by obtaining an additional label for majority voting. All in all, each student produced a minimum of 300 annotated sentences. While the students had the option to collect data in any language, all of them, apart from two, collected and annotated the initial data in English. The two other students who collected data in different languages chose their native languages: Telugu, and Hindi.",
+        "bbox": [
+            507,
+            102,
+            884,
+            374
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To collect the data, the students were allowed to do anything they wanted. They ended up utilizing diverse techniques that range from targeted web scraping to generating sentences with the assistance of AI tools such as, ChatGPT (Radford et al., 2019). We can broadly categorize the sources of data into three categories: AI tools (such as ChatGPT and ChatSonic), online news platforms (including Online Articles, NBC, CNN, BBC, and NYTimes), and social media platforms (such as Twitter and Reddit). Students have used a combination of two or more categories to collect their data. Around  $77\\%$  of students used AI tools,  $14.8\\%$  relied on social media platforms, and  $67.9\\%$  used online news platforms for data collection purposes. It is important to note that, AI was only used by the students in the first step of data collection. This shows how artificial intelligence (AI) eases the process of collecting relevant, topic-specific text. The process of data validation and labeling was entirely done by human annotators.",
+        "bbox": [
+            507,
+            376,
+            884,
+            712
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In the end, we ended up with a total of 17,520 sentences from the combined student training corpus of 300 sentences each, eliminating the occasional duplicate instances. The dataset has a generally substantial inter-annotator agreement, with a Cohen's  $\\kappa$  (Cohen, 1960) coefficient of 0.61.",
+        "bbox": [
+            507,
+            715,
+            884,
+            810
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To further contextualize this, we note that the inter-annotator agreement of the MFC (as detailed in the paper) is assessed using Krippendorff's  $\\alpha$  (Krippendorff, 2011), with respective values of 0.08 and 0.20 for the domains of same-sex marriage and im",
+        "bbox": [
+            507,
+            813,
+            884,
+            894
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_footnote",
+        "text": "2We are releasing these data with the students' consent.",
+        "bbox": [
+            134,
+            906,
+            472,
+            920
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_footnote",
+        "text": "These evaluation sets were based on the MFC test sets.",
+        "bbox": [
+            529,
+            906,
+            862,
+            919
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "4158",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "migration. SNFC (our dataset) combines sentences from both of these domains and the Krippendorff's  $\\alpha$  value for SNFC stands at 0.103 which is similar to the one of MFC. Given that this is a 15-way classification task, we believe the inter-annotator agreement for SNFC is not particularly low for such a nuanced task.",
+        "bbox": [
+            112,
+            84,
+            487,
+            197
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Multilinguality To benchmark media framing beyond English our first step is to simply translate the original MFC dataset into other languages. We use machine translation<sup>4</sup> to translate all sentences of the MFC corpus into 12 typologically diverse languages, namely Bengali, German, Greek, Italian, Turkish, Nepali, Hindi, Portuguese, Telugu, Russian, Swahili, and Mandarin Chinese.",
+        "bbox": [
+            110,
+            206,
+            487,
+            332
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "While the primary reason for this process is the ability to benchmark the task on other languages (as well as the inability to collect annotated test sets in all of these languages - see also §3), this simple data augmentation technique is also a reasonable way to also obtain training data in other languages. Hence, we perform this translation both on the training and the dev/test portions of the dataset, and combine all languages to form the multilingual version of the dataset.",
+        "bbox": [
+            112,
+            335,
+            487,
+            495
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Lastly, the same translation models were used to augment our crowd-sourced SNFC dataset to cover all of the above-mentioned languages.",
+        "bbox": [
+            112,
+            497,
+            487,
+            545
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We have studied the quality of the translation through human assessment. For each language, we took 100 translations from English and had them reviewed by bilingual speakers who scored the translations on a scale from 1 to 10 based on accuracy and clarity. For this evaluation, we used four languages: Bengali, Greek, Hindi, and Nepali. From the average rating for each language pair (See Table 1), we observe that the average rating is higher for higher resourced languages like Greek and Hindi. On the other hand, Nepali, being the only lower resourced language, has a lower rating of 4.72 out of 10, suggesting that perhaps Nepali results should be taken with a grain of salt, as the reason for general poor performance is likely to be the low quality of the translations.",
+        "bbox": [
+            112,
+            546,
+            487,
+            802
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We have also further performed quality estimation over all translations by calculating the CometKiwi score (Rei et al., 2023) of the translations. Note that we resort to automatic quality estimation since we do not have access to reference translations. The overall score of  $76.05\\%$  is",
+        "bbox": [
+            112,
+            804,
+            487,
+            898
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/b5c109e70e3bd09109c59c34c67eed0eba8bb94e249450ba1a79fc57cd8573cb.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Language Pair</td><td>Rating (%)</td></tr><tr><td>English-Bengali</td><td>61.2</td></tr><tr><td>English-Greek</td><td>73.4</td></tr><tr><td>English-Hindi</td><td>77.4</td></tr><tr><td>English-Nepali</td><td>47.2</td></tr><tr><td>Comet Score (All languages)</td><td>76.05</td></tr></table>",
+        "bbox": [
+            522,
+            80,
+            870,
+            202
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Average rating for Human Evaluation of the Automatic Translation Quality",
+        "bbox": [
+            507,
+            211,
+            880,
+            240
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "in line with our human evaluation over the sample, and suggests that automatic translations are largely reliable in our dataset. The higher scores for the high resource languages of the human-evaluation and CometKiwi (see Appendix C for a breakdown by language) indicate that automatic translations can be a reasonable alternative to gathering large quantities of high quality multilingual data for the framing task.",
+        "bbox": [
+            507,
+            242,
+            882,
+            387
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Novel Test Set While the automatic translation of the MFC benchmark is a reasonable start for our multilingual exploration, it does not come without drawbacks: the provided text, regardless of the language, is only relevant to the USA cultural context.",
+        "bbox": [
+            507,
+            401,
+            882,
+            481
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To even better benchmark the quality of framing analysis systems on different language and cultural contexts, we create a pair of novel test sets in (Bangladesh) Bengali and (Brazilian) Portuguese. The news articles used in this test set were sourced from reputable newspapers in Bangladesh and Brazil, aligning with the chosen domains of immigration and same-sex marriage.",
+        "bbox": [
+            507,
+            482,
+            882,
+            611
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Each test set is comprised of 10 news articles for each language. The annotators were native speakers of the languages and they adhered closely to the definitions provided by the authors (Table 5), ensuring consistency with the labels found in the MFC.",
+        "bbox": [
+            507,
+            612,
+            882,
+            706
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Figure 2 shows the label distribution for the MFC and the novel test set, listing the number of sentences per frame in each language. In the case of Bengali, the news articles predominantly focus on the immigration domain, reflecting the cultural disparities between Brazil and Bangladesh. Specifically, the test set emphasizes the economic and lifestyle aspects of immigration (Bengali), while also delving into the legal and policy-making dimensions of the domain (Portuguese).",
+        "bbox": [
+            507,
+            709,
+            882,
+            871
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "It is of note that the two benchmarks, despite being rather small, still show interesting differences in terms of their label distribution. For example, the",
+        "bbox": [
+            507,
+            873,
+            882,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "4 Google Translate, specifically.",
+        "bbox": [
+            134,
+            906,
+            327,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "4159",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/bb5830acf1a76884af095a52ca26b276aaeb48bb2acfa5bcee9071befdb2be4e.jpg",
+        "image_caption": [
+            "Figure 2: The label distributions of the MFC and our new Bengali and Portuguese test sets. Note that they differ significantly."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            159,
+            70,
+            843,
+            288
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "most common label on the Bengali set is \"External Regulation and Reputation\", which is the least common one in the Portuguese one. And the reverse is the case for the \"Cultural Identity\" label which is the most common in Portuguese and least common in Bengali. Another interesting observation is that the Bengali test set contains more data labeled as \"Other\" compared to the other two languages. Upon analyzing the data with the help of a native speaker, we found that most of the Bangladeshi articles emphasize a lot on reporting information in the form of dates and numbers, rather than offering opinions on the issues.",
+        "bbox": [
+            112,
+            347,
+            489,
+            556
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4 Framing Analysis System and Results",
+        "text_level": 1,
+        "bbox": [
+            112,
+            570,
+            473,
+            586
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Experimental Setup We approach the task as a multilabel classification problem (Tsoumakas and Katakis, 2007), leveraging the pretrained RoBERTa (Zhuang et al., 2021) language model, similar to the SOTA approach employed by Khanehzar et al. (2019). For all models we set the maximum sequence length to 256, with a batch size of 16, and train using a learning rate of  $10^{-5}$ . To expand to more languages, we employ the multilingual XLM-RoBERTa model (Conneau et al., 2020). Throughout all experiments, we use the base model size.",
+        "bbox": [
+            112,
+            596,
+            489,
+            788
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We first report results with models exclusively trained on MFC, and SNFC datasets, as well as their concatenation. To investigate a more data-scarce scenario, we also compiled a smaller sample consisting of about  $10\\%$  of the original MFC,",
+        "bbox": [
+            112,
+            791,
+            489,
+            873
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/855c4238f97acdb93cb20c84d9c426267350db4cbc63571548a5069af02300db.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Tr. Data</td><td>#Sentences</td><td>Accuracy</td></tr><tr><td colspan=\"3\">Baselines</td></tr><tr><td>MFC</td><td>9739</td><td>69.52</td></tr><tr><td>MFC10</td><td>1125</td><td>57.45</td></tr><tr><td colspan=\"3\">including crowd-sourced data</td></tr><tr><td>SNFC</td><td>17520</td><td>54.37</td></tr><tr><td>SNFC50</td><td>8760</td><td>54.7</td></tr><tr><td>MFC+SNFC</td><td>27260</td><td>72.07</td></tr><tr><td>MFC+SNFC50</td><td>18499</td><td>72.89</td></tr><tr><td>MFC10+SNFC</td><td>18645</td><td>64.75</td></tr><tr><td>MFC10+SNFC50</td><td>9885</td><td>62.05</td></tr><tr><td colspan=\"3\">filtered crowd-sourced data</td></tr><tr><td>MaSNFC</td><td>5182</td><td>48.77</td></tr><tr><td>MFC+MaSNFC</td><td>14922</td><td>73.22</td></tr><tr><td>MFC10+MaSNFC</td><td>6307</td><td>60.94</td></tr></table>",
+        "bbox": [
+            531,
+            344,
+            862,
+            615
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Mean Accuracy Scores on the MFC evaluation set for RoBERTa models trained on English Datasets. # stands for \"number of\".",
+        "bbox": [
+            507,
+            625,
+            882,
+            667
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "named MFC10, ensuring all 15 target labels are included. Beyond the single-dataset baselines, we combine the expert-annotated MFC and MFC10 with our crowd-sourced SNFC. To further study the effect of the size of the SNFC, we have experimented with SNFC50, a randomized halved subset of the original SNFC that is more closer to the MFC in size.",
+        "bbox": [
+            507,
+            669,
+            882,
+            781
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "English Results and Discussion We first establish the usefulness of our crowdsourced data, by focusing on the performance on the original test set of the English MFC dataset (using the monolingual RoBERTa model). Results are presented in Table 2.",
+        "bbox": [
+            507,
+            791,
+            884,
+            872
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "First, it is worth pointing out that relying solely on crowd-sourced data is not promising: the SNFC-only training underperforms both the MFC-only set",
+        "bbox": [
+            507,
+            873,
+            882,
+            921
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "5Appendix 8 and 9 also provides results with the BERT and mBERT (Devlin et al., 2019) models (but RoBERTa and XLM-R consistently outperformed BERT and mBERT.",
+        "bbox": [
+            112,
+            882,
+            487,
+            921
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "4160",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "ting, as well as the MFC10-only setting, which has only around  $10\\%$  of the training data size!",
+        "bbox": [
+            112,
+            84,
+            485,
+            116
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "However, combining the expert-annotated data with the crowd-sourced ones yields significant improvements over the expert-only baselines, as MFC+SNFC yields an extra 2.5 accuracy points over MFC (72% vs 69.5%). The improvement is even larger (more than 7 accuracy points) in the resource-restricted MFC10 scenario. The accuracy remains consistent both with SNFC50 alone and when combined with MFC, as MFC+SNFC50 and MFC+SNFC yield similar results, indicating that performance gains are not merely due to larger data volume.",
+        "bbox": [
+            110,
+            118,
+            487,
+            294
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Filtering of Crowdsourced Data Given the potential for noise in any crowd-sourced dataset, we explore a simple filtering technique to sample more high-quality crowd-sourced. In particular, we obtain sentence-level representations for each sentence, and select only the SNFC instances that exhibit more than  $85\\%$  cosine similarity with any MFC instance. Effectively, we select SNFC sentences that are most similar to MFC ones. We refer to this sample as MFC-aligned SNFC (MaSNFC).",
+        "bbox": [
+            110,
+            309,
+            489,
+            469
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Results with this (almost 3x smaller) sample are more encouraging (Table 2): combining MaSNFC with MFC yields our best model with an accuracy of 73.22. In the data-scarce scenario of MFC10, adding MaSNFC is again beneficial, but including the whole unfiltered SNFC is even better.",
+        "bbox": [
+            112,
+            470,
+            487,
+            565
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "These findings underline the promise of crowdsourcing for collecting a high volume of (somewhat) lower quality data. The performance improvement for the MaSNFC+MFC shows promise for the combination of low-volume high-quality along with a higher-volume of lower-quality data. This approach effectively balances the depth and breadth of the dataset, leveraging the strengths of both data types.",
+        "bbox": [
+            112,
+            569,
+            487,
+            714
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Multilingual Results and Discussion For the first part of our multilingual experiments, we employ a translate-train and translate-test scenario. All of the dataset samples introduced above were translated to all 12 evaluation languages, and we now replicate the same experimental setups as above, the only difference being that we will use a multilingual LM (XLM-R instead of RoBERTa). All results are presented in Table 3 (which presents the average accuracy across the 12 languages for mMFC, as well as performance on our novel Bengali and Portuguese benchmark).",
+        "bbox": [
+            112,
+            728,
+            489,
+            920
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/cb45c43fb008f886ebe368ffb14c875bf314a3a207364c0c552b434a25378089.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Tr. Data</td><td>mMFC</td><td>BENGALI</td><td>PORTUGUESE</td></tr><tr><td colspan=\"4\">Zero-shot (only English train)</td></tr><tr><td>MFC</td><td>28.13</td><td>25.44</td><td>28.28</td></tr><tr><td colspan=\"4\">Baselines (translate-train)</td></tr><tr><td>MFC</td><td>44.99</td><td>25.88</td><td>33.61</td></tr><tr><td>MFC10</td><td>28.64</td><td>23.68</td><td>27.87</td></tr><tr><td colspan=\"4\">+ crowd-sourced (translate-train)</td></tr><tr><td>SNFC</td><td>28.04</td><td>25.44</td><td>23.77</td></tr><tr><td>MFC+SNFC</td><td>44.07</td><td>26.31</td><td>31.56</td></tr><tr><td>MFC10+SNFC</td><td>33.11</td><td>32.02</td><td>26.62</td></tr><tr><td colspan=\"4\">+ filtered crowd-sourced (translate-train)</td></tr><tr><td>MaSNFC</td><td>27.55</td><td>16.67</td><td>15.98</td></tr><tr><td>MFC+MaSNFC</td><td>45.73</td><td>28.07</td><td>33.61</td></tr><tr><td>MFC10+MaSNFC</td><td>32.56</td><td>24.56</td><td>26.64</td></tr></table>",
+        "bbox": [
+            519,
+            82,
+            872,
+            336
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Mean Accuracy Scores on the MFC evaluation set and Novel Multilingual Test Set for XLM-R models trained on Multilingual Datasets. The best scores have been highlighted.",
+        "bbox": [
+            507,
+            346,
+            882,
+            404
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "First of all, we show that relying on zero-shot cross-lingual transfer, without employing the translate-train technique is not a competitive baseline. The translated MFC baseline is competitive on average, but as we discuss below it performs quite inequitably across languages. As before, combining expert annotated data with filtered crowdsourced ones (MFC+MaSNFC) is best. Our findings from the monolingual experiments generally hold in the multilingual ones.",
+        "bbox": [
+            507,
+            419,
+            882,
+            580
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In the Bengali test set, the inclusion of all crowdsourced data improves upon the baseline by a small margin. The improvement from filtered crowdsourced data is more modest. However, it is interesting that the best performance is obtained when using fewer expert annotations (MFC10+SNFC), improving by almost 6 percentage points over the baseline! We hypothesize that using the whole MFC dataset overfits the US context – but we leave this analysis for future work. In the Portuguese test set, we observe generally similar patterns as in the mMFC, with the exception that we do not observe any improvement from the crowd-sourced data. We leave a further investigation for future work.",
+        "bbox": [
+            507,
+            582,
+            882,
+            806
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We note that the accuracies for the Bengali and Portuguese test sets are significantly lower than those of the English MFC and the mMFC test sets. We suspect that the training data, being automatic translations, may not capture the nuances of the original news articles. Second, the domain shift due to cultural context differences between training and test",
+        "bbox": [
+            507,
+            808,
+            882,
+            920
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "4161",
+        "bbox": [
+            480,
+            928,
+            517,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/89f04374dd139fe0c5bae757792230560d2a822da3a6392020b9d15c164d7e49.jpg",
+        "image_caption": [
+            "Figure 3: The best model performs very inequitably across languages on mMFC. The highest accuracy is in English (72.1%) followed by Italian and German, while other languages from non-western countries (e.g. Bengali, Hindi, Chinese, and others) have much lower performance (under 30%)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            200,
+            71,
+            796,
+            275
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "may play a significant role. To improve the scores further, it may be necessary to obtain original news articles from diverse culturally distinct sources in different languages.",
+        "bbox": [
+            112,
+            346,
+            487,
+            411
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "mMFC Breakdown per Language We further analyse the per-language performance of our best-performing model on mMFC (see Figure 3). English accuracy (72.1) is en par with the monolingual setting (73.2), and German, Italian, Swedish, and Turkish also yield accuracies higher than  $64\\%$ . But for other languages the model performs much worse, including high-resource ones like Greek  $(31.5\\%)$ , Russian  $(28\\%)$ , and Chinese  $(25.5\\%)$ . While translation errors may play a role here, we are confident that they are not enough to explain such a large discrepancy. For example, while Nepali has admittedly low-quality translations (see previous discussion), Hindi, Greek, and Chinese certainly have translations of fairly high quality and yet they fall in the same low performance ballpark. We suspect that this gap may only be bridged through data collection (either expert- or crowd-annotated) in the appropriate languages and cultural contexts.",
+        "bbox": [
+            115,
+            424,
+            489,
+            745
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Error Analysis We analyzed the errors using a confusion matrix for our best-performing model MFC+MaSNFC on the mMFC evaluation set, as shown in Figure 4. The heat-map reveals that out of 15 labels, 9 achieve the majority of instances correctly. Specifically, the labels 'Political' and 'Legality, Constitutionality, Jurisdiction' have the highest number of instances predicted correctly. However, when the model makes incorrect predictions, the errors are mainly categorized into the 'Political'",
+        "bbox": [
+            112,
+            760,
+            489,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "and 'Legality, Constitutionality, Jurisdiction' labels. This led us to suspect a potential data imbalance in our training model. Further examination of the data confirmed that these two labels indeed have a majority of instances in the training set, leading to the tendency to predict these labels when uncertain.",
+        "bbox": [
+            507,
+            346,
+            884,
+            441
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "One could also further argue that these two labels are quite close semantically and hence their confusion is perhaps expected. We have examined the original data from MFC for the immigration and same-sex issues, which were used to train our baseline model. This dataset indeed shows a skewed distribution with a disproportionate number of instances falling under these two labels. This suggests that US-based news articles covering these domains inherently tend to fall in these two categories. Given the domain, we deduce that such an imbalance in label distribution might be a common trend in news articles from other countries as well. This assumption can be further validated in our novel test sets derived from Bangladesh and Brazil, which also reveal a similar inclination towards certain labels, as discussed in the previous section.",
+        "bbox": [
+            507,
+            443,
+            884,
+            731
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5 Generative Language Models",
+        "text_level": 1,
+        "bbox": [
+            507,
+            747,
+            798,
+            765
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "LLMs like GPT-4 (OpenAI, 2023), Falcon (Penedo et al., 2023), and LLaMA (Touvron et al., 2023), are trained on vast amounts of text and have shown immense promise in a variety of NLP tasks. Their broad knowledge base qualifies them as potential tools for framing analysis. In this study, we have also explored three of these models, particularly the open-sourced ones: Mistral, LLaMA-2, and Falcon.",
+        "bbox": [
+            507,
+            776,
+            884,
+            919
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "4162",
+        "bbox": [
+            480,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/51f87766689f38ea67ee7eda04b48002e2204b35fdcfcb3a104deea857e9c7cc.jpg",
+        "image_caption": [
+            "Figure 4: Confusion matrix for the best model's prediction for the mMFC Test set."
+        ],
+        "image_footnote": [
+            "Economic: E  \nCapacity and Resources: CaR  \nMorality: M  \nFairness and Equality: FaE  \nLegality, Constitutionality, Jurisdiction: LCJ  \nPolicy Prescription and Evaluation: PPaE  \nCrime and Punishment: CaP  \nSecurity and Defense: SaD  \nHealth and Safety: HaS  \nQuality of Life: QoL  \nCultural Identity: CI  \nPublic Sentiment: PS  \nPolitical: P  \nExternal Regulation and Reputation: ERaR  \nOther: O"
+        ],
+        "bbox": [
+            121,
+            71,
+            606,
+            382
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/63e396a348a5ee71d37a51110b439fa870d02b21eedeef901b49e37eda8bb843.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>Accuracy (%)</td></tr><tr><td>Falcon-40b-instruct</td><td>22.95</td></tr><tr><td>Mistral-7B-Instruct-v0.1</td><td>35.33</td></tr><tr><td>Llama2-chat-70B</td><td>22.22</td></tr></table>",
+        "bbox": [
+            127,
+            419,
+            475,
+            500
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 4: Exact Match accuracy of the LLMs. The highest accuracy (35%, bolded) is significantly worse than the task finetuned RoBERTa model's performance (73.22%).",
+        "bbox": [
+            112,
+            508,
+            487,
+            567
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Experimental Setting The instruction presents the framing task as a multiple choice question with 15 options and we have curated the instruction to include the definitions of all the labels, similar to the ones the students have used to annotate the SNFC. The instruction we use is given in Appendix E. We conduct all experiments in the zero shot setting, to assess the potential of LLMs to generalize and apply their knowledge effectively without task-specific training. The experiments were run on the English only test set (MFC-test) to ensure comparability with other task-finetuned models previously evaluated on the same test set.",
+        "bbox": [
+            112,
+            583,
+            489,
+            791
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Results and Discussion The results (see Table 4) show the exact match accuracy of different LLMs on the MFC-test dataset. The performance of Llama2-chat-70B aligns closely with that of Falcon-40b-instruct, and Mistral-7B-Instruct-v0.1 outperformed them significantly showing that the sheer size of a model does not necessarily equate",
+        "bbox": [
+            112,
+            808,
+            489,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "to better performance.",
+        "bbox": [
+            507,
+            422,
+            677,
+            437
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Interestingly, the best performance was achieved by employing smaller, task-finetuned models, with RoBERTa achieving an exact match accuracy of  $73.22\\%$ . This significantly surpasses the highest result for general LLMs, as their best performance is at  $35.33\\%$ , observed with Mistral-7B-Instructv0.1. This difference in performance highlights the importance of task-specific fine-tuning on model efficacy. The finetuning process allows models like RoBERTa to adapt their parameters more closely to the nuances of the specific task, resulting in a more precise understanding and response generation compared to models that rely solely on broad, generalized training. The results also suggests that there is a trade-off between model size and specialized training. While larger models have a vast knowledge base, they are not always effective in applying this knowledge to specific tasks without fine-tuning.",
+        "bbox": [
+            507,
+            439,
+            884,
+            745
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Error Analysis The LLMs exhibit a range of errors in predicting the correct frames for the provided texts (See Table 10). These errors include spelling mistakes, overgeneralization, assigning multiple labels where only one is appropriate, and misinterpretation. Generally, the models struggle with adhering to instructions, such as inventing new frames rather than selecting from the provided list (External Regulatory and Renown). Additionally, a common issue among all three of the models is",
+        "bbox": [
+            507,
+            760,
+            884,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "4163",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "their failure to introduce their answers concisely as instructed. Contrary to the clear direction to reply only with the label name, they begin responses with phrases like 'The most suitable frame is...'.",
+        "bbox": [
+            112,
+            84,
+            487,
+            148
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "The Mistral 7B model achieves a higher accuracy rate compared to the other two model; however, it often adds additional commentary to its responses. The LLaMA-2 70B model's predictions are inconsistent, notably when it replaces 'External Regulation and Reputation' with 'External Regulatory and Renown', demonstrating a tendency towards misrepresentation. The Falcon 40B sometimes accurately identifies the frame but fails to use the exact label name, responding with 'Economical' instead of 'Economic'.",
+        "bbox": [
+            115,
+            149,
+            489,
+            324
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Since the models have the tendency to predict labels with spelling errors and synonymous labels, we have employed different techniques to measure the accuracy of these models to ensure a true reflection of the system's performance. To derive the correct label names from synonymous words and to overlook spelling mistakes, we employed the Fast-Text (Joulin et al., 2016) and Edit Distance (Levenshtein et al., 1966) algorithms. These were used to determine the textual similarity between the models' predictions and the 15 labels they were intended to predict.",
+        "bbox": [
+            115,
+            326,
+            489,
+            519
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "6 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            112,
+            533,
+            247,
+            548
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In conclusion, our study emphasizes the importance of data quality and language diversity in multilingual framing analysis. Combining the Media Frames Corpus (MFC) with the Student-Sourced Noisy Frames Corpus (SNFC) yields significant improvements, highlighting the value of larger datasets despite the annotation quality potentially being lower. However, lower accuracy in multilingual experiments indicates the need for accurate translations and culturally diverse training data to improve multilingual framing analysis. Last, the sub-par performance of LLMs showcases a future research direction towards task-specific finetuning of the LLMs.",
+        "bbox": [
+            112,
+            560,
+            489,
+            783
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            112,
+            797,
+            220,
+            813
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "The main limitation of this study is that it relies on automated translation via Google Translator to introduce multilinguality to the task. It is well known that the translations conducted by Google Translator may not achieve the same level of quality as authentic translations. Moreover, for lower",
+        "bbox": [
+            112,
+            825,
+            489,
+            921
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "resource languages such as Nepali and Swahili, the translations obtained from Google Translator may not fully capture the nuances and characteristics as well as it probably can if translated to higher-resource languages as German or Greek. Additionally, since the MFC dataset primarily consists of US news sources, the translations into different languages does not adequately reflect the biases and perspectives surrounding a specific political issue in different countries. We attempt to mitigate this limitation with our new Bengali and Portuguese test sets. Collecting more data from different countries in different languages will eventually address this limitation, but we leave this large-scale undertaking for the future.",
+        "bbox": [
+            507,
+            84,
+            885,
+            325
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            509,
+            338,
+            680,
+            355
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We are thankful to the anonymous reviewers for their useful feedback, as well as to the students of the GMU CS 678 course and the annotators for the Bengali and Portuguese test set who majorly contributed in the creation of our crowdsourced dataset. This project was supported by the National Science Foundation under grant IIS-2327143. This project was also supported by resources provided by the Office of Research Computing at George Mason University (https://orc.gmu.edu) and funded in part by grants from the National Science Foundation (Awards Number 1625039 and 2018631).",
+        "bbox": [
+            507,
+            365,
+            885,
+            557
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            510,
+            586,
+            608,
+            601
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Afra Feyza Akyurek, Lei Guo, Randa I. Elanwar, Prakash Ishwar, Margrit Betke, and Derry Tanti Wijaya. 2020. Multi-label and multilingual news framing analysis. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 8614-8624. Association for Computational Linguistics.",
+            "Mohammad Ali and Naeemul Hassan. 2022. A survey of computational framing analysis approaches. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, EMNLP 2022, Abu Dhabi, United Arab Emirates, December 7-11, 2022, pages 9335-9348. Association for Computational Linguistics.",
+            "Amber E Boydstun, Dallas Card, Justin Gross, Paul Resnick, and Noah A Smith. 2014. Tracking the development of media frames within and across policy issues.",
+            "Bjorn Burscher, Rens Vliegenthart, and Claes H de Vreese. 2016. Frames beyond words: Applying cluster and sentiment analysis to news coverage of the"
+        ],
+        "bbox": [
+            509,
+            609,
+            885,
+            921
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "4164",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "nuclear power issue. Social Science Computer Review, 34(5):530-545.",
+            "Dallas Card, Amber E. Boydstun, Justin H. Gross, Philip Resnik, and Noah A. Smith. 2015. The media frames corpus: Annotations of frames across issues. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing of the Asian Federation of Natural Language Processing, ACL 2015, July 26-31, 2015, Beijing, China, Volume 2: Short Papers, pages 438-444. The Association for Computer Linguistics.",
+            "Jacob Cohen. 1960. A coefficient of agreement for nominal scales. Educational and psychological measurement, 20(1):37-46.",
+            "Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettle-moyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 8440-8451. Association for Computational Linguistics.",
+            "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 4171-4186. Association for Computational Linguistics.",
+            "Paul DiMaggio, Manish Nag, and David Blei. 2013. Exploiting affinities between topic modeling and the sociological perspective on culture: Application to newspaper coverage of us government arts funding. Poetics, 41(6):570-606.",
+            "Robert M Entman. 1993. Framing: Toward clarification of a fractured paradigm. Journal of communication, 43(4):51-58.",
+            "Anjalie Field, Doron Kliger, Shuly Wintner, Jennifer Pan, Dan Jurafsky, and Yulia Tsvetkov. 2018. Framing and agenda-setting in russian news: a computational analysis of intricate political strategies. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Brussels, Belgium, October 31 - November 4, 2018, pages 3570-3580. Association for Computational Linguistics.",
+            "Mokter Hossain and Ilkka Kauranen. 2015. Crowdsourcing: a comprehensive literature review. Strategic Outsourcing: An International Journal, 8(1):2-22.",
+            "Jeff Howe. 2006. The rise of crowdsourcing, wired. http://www.wired.com/wired/archive/14.06/crowds.html."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Armand Joulin, Edouard Grave, Piotr Bojanowski, Matthijs Douze, Hérve Jégou, and Tomas Mikolov. 2016. Fasttext.zip: Compressing text classification models. arXiv preprint arXiv:1612.03651.",
+            "Shima Khanehzar, Andrew Turpin, and Gosia Mikolajczak. 2019. Modeling political framing across policy issues and contexts. In Proceedings of the The 17th Annual Workshop of the Australasian Language Technology Association, ALTA 2019, Sydney, Australia, December 4-6, 2019, pages 61-66. Australasian Language Technology Association.",
+            "Klaus Krippendorff. 2011. Computing krippendorff's alpha-reliability.",
+            "Haewoon Kwak, Jisun An, and Yong-Yeol Ahn. 2020. A systematic media frame analysis of 1.5 million new york times articles from 2000 to 2017. In WebSci '20: 12th ACM Conference on Web Science, Southampton, UK, July 6-10, 2020, pages 305-314. ACM.",
+            "Vladimir I Levenshtein et al. 1966. Binary codes capable of correcting deletions, insertions, and reversals. In Soviet physics doklady, volume 10, pages 707-710. Soviet Union.",
+            "Siyi Liu, Lei Guo, Kate Mays, Margrit Betke, and Derry Tanti Wijaya. 2019. Detecting frames in news headlines and its application to analyzing news framing trends surrounding U.S. gun violence. In Proceedings of the 23rd Conference on Computational Natural Language Learning (CoNLL), pages 504-514, Hong Kong, China. Association for Computational Linguistics.",
+            "Julia Mendelsohn, Ceren Budak, and David Jurgens. 2021. Modeling framing in immigration discourse on social media. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2219-2263, Online. Association for Computational Linguistics.",
+            "Nona Naderi and Graeme Hirst. 2017. Classifying frames at the sentence level in news articles. In Proceedings of the International Conference Recent Advances in Natural Language Processing, RANLP 2017, pages 536-542, Varna, Bulgaria. INCOMA Ltd.",
+            "Viet-An Nguyen. 2015. Guided probabilistic topic models for agenda-setting and framing. Ph.D. thesis, University of Maryland, College Park.",
+            "OpenAI. 2023. Gpt-4 technical report. ArXiv, abs/2303.08774.",
+            "Guilherme Penedo, Quentin Malartic, Daniel Hesslow, Ruxandra Cojocaru, Alessandro Cappelli, Hamza Alobeidli, Baptiste Pannier, Ebtesam Almazrouei, and Julien Launay. 2023. The refinedweb dataset for falcon llm: Outperforming curated corpora with web data, and web data only."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            919
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "4165",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9.",
+            "Ricardo Rei, Nuno M Guerreiro, José Pombal, Daan van Stigt, Marcos Treviso, Luisa Coheur, José GC de Souza, and André FT Martins. 2023. Scaling up cometkiwi: Unbabel-ist 2023 submission for the quality estimation shared task. arXiv preprint arXiv:2309.11925.",
+            "Margaret E Roberts, Brandon M Stewart, Dustin Tingley, Christopher Lucas, Jetson Leder-Luis, Shana Kushner Gadarian, Bethany Albertson, and David G Rand. 2014. Structural topic models for open-ended survey responses. American journal of political science, 58(4):1064-1082.",
+            "Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. 2023. Llama 2: Open foundation and finetuned chat models.",
+            "Grigorios Tsoumakas and Ioannis Katakis. 2007. Multilabel classification: An overview. International Journal of Data Warehousing and Mining (IJDWM), 3(3):1-13.",
+            "Yuxiang Zhao and Qinghua Zhu. 2014. Evaluation on crowdsourcing research: Current status and future direction. Information systems frontiers, 16:417-434.",
+            "Liu Zhuang, Lin Wayne, Shi Ya, and Zhao Jun. 2021. A robustly optimized BERT pre-training approach with post-training. In Proceedings of the 20th Chinese National Conference on Computational Linguistics, pages 1218-1227, Huhhot, China. Chinese Information Processing Society of China."
+        ],
+        "bbox": [
+            115,
+            85,
+            485,
+            826
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "4166",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A Annotation Schema",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            324,
+            99
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "We used a crowdsourcing approach with the help of non-expert annotators to create our training corpus, simplifying the process compared to the traditional method of hand-annotating by expert linguists and social science scholars, which is both expensive and inefficient. We collected data for the corpus in collaboration with graduate students whose task was to gather 150 sentences each, in various languages, from news articles related to the domains of immigration and same-sex marriage. These sentences were then annotated using the 15 framing dimensions established in the study (Boydstun et al., 2014), which are globally accepted, shown in Table 5.",
+        "bbox": [
+            112,
+            109,
+            492,
+            335
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/28d96433634512a0e8b149def9c86c6c4ce49f2d186b4eb9dc281e1971b0926c.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Frames</td><td>Definitions</td></tr><tr><td>Economic</td><td>The financial consequences and economic implications of the matter on various levels (person, family, community or broader economy).</td></tr><tr><td>Capacity and Resources</td><td>The presence or absence of various resources(physical, geographic, human, and financial) and the ability of existing systems.</td></tr><tr><td>Morality</td><td>Perspectives, policy objectives, or actions driven by religious principles, duties, ethics, or social responsibilities.</td></tr><tr><td>Fairness and Equality</td><td>The balance or distribution of laws, rights, and resources among individuals or groups.</td></tr><tr><td>Legality, Constitutionality, Jurisdiction</td><td>Discusses rights, freedoms and authority of individuals, corporations, and government.</td></tr><tr><td>Policy Prescription and Evaluation</td><td>Specific policies proposed to address identified issues and the assessment of policy effectiveness.</td></tr><tr><td>Crime and Punishment</td><td>Effectiveness and implications of laws and their enforcement.</td></tr><tr><td>Security and Defense</td><td>Actions or calls to action aimed at protecting individuals, groups, or nations from potential threats to their well-being.</td></tr><tr><td>Health and Safety</td><td>Access to healthcare, health outcomes, disease, sanitation, mental health, violence prevention, infrastructure safety, and public health.</td></tr><tr><td>Quality of life</td><td>Threats and opportunities for the individual&#x27;s wealth, happiness and well being.</td></tr><tr><td>Cultural Identity</td><td>Traditions, customs or values of a social group in relation to a policy issue.</td></tr><tr><td>Public Sentiment</td><td>References of attitudes and opinions of the general public, including polling and demographics.</td></tr><tr><td>Political</td><td>Political considerations, actions, efforts, stances, and partisan, bipartisan, or lobbying activities related to an issue.</td></tr><tr><td>External Regulation and Rep- ution</td><td>The external relations of nations or groups, trade agreements, policy outcomes, and external perceptions or consequences.</td></tr><tr><td>Other</td><td>Frames that don&#x27;t fit into the categories above.</td></tr></table>",
+        "bbox": [
+            115,
+            361,
+            850,
+            878
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Table 5: Frames and their definitions as outlined by Policy Frames Codebook (PFC, Boydstun et al. (2014)). This codebook was given to the students as annotation schema.",
+        "bbox": [
+            114,
+            887,
+            882,
+            917
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "4167",
+        "bbox": [
+            480,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/fead6e7d0acfc4b2c1d4638e2c6040618abdd954d525efaaa7f64c673e40ed7b.jpg",
+        "table_caption": [
+            "B Novel Bengali and Portuguese Test Set Statistic"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Number of sentences</td><td>Bengali</td><td>Portuguese</td></tr><tr><td>Economic</td><td>36</td><td>20</td></tr><tr><td>Capacity and Resources</td><td>3</td><td>19</td></tr><tr><td>Morality</td><td>4</td><td>13</td></tr><tr><td>Fairness and Equality</td><td>13</td><td>23</td></tr><tr><td>Legality Constitutional-ity Jurisdiction</td><td>12</td><td>25</td></tr><tr><td>Policy Prescription and Evaluation</td><td>13</td><td>24</td></tr><tr><td>Crime and Punishment</td><td>11</td><td>3</td></tr><tr><td>Security and Defence</td><td>5</td><td>23</td></tr><tr><td>Health and Safety</td><td>14</td><td>9</td></tr><tr><td>Quality of Life</td><td>33</td><td>15</td></tr><tr><td>Cultural Identity</td><td>1</td><td>32</td></tr><tr><td>Public Sentiment</td><td>5</td><td>24</td></tr><tr><td>Political</td><td>3</td><td>10</td></tr><tr><td>External Regulation and Reputation</td><td>41</td><td>1</td></tr><tr><td>Other</td><td>34</td><td>3</td></tr><tr><td>Total</td><td>228</td><td>244</td></tr></table>",
+        "bbox": [
+            119,
+            134,
+            482,
+            479
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Table 6: Number of texts per frame per language",
+        "bbox": [
+            134,
+            489,
+            465,
+            504
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "The distribution of labels in the Bengali and Portuguese test sets (see Table 6) reveals intriguing domain affinity. In the case of Bengali, the news articles predominantly focus on the immigration domain, reflecting the cultural disparities between Brazil and Bangladesh. Specifically, the test set emphasizes the economic and lifestyle aspects of immigration (Bengali), while also delving into the legal and policy-making dimensions of the domain (Portuguese).",
+        "bbox": [
+            112,
+            523,
+            489,
+            683
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "4168",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "C Assessing Translation Quality",
+        "text_level": 1,
+        "bbox": [
+            115,
+            83,
+            410,
+            101
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "table",
+        "img_path": "images/bcb6ffc20212efa8161935c35bc06552b83f56050d94b1f5c0f1d914b396ef58.jpg",
+        "table_caption": [
+            "Table 7 shows the breakdown of the comet score per language."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Language Pair</td><td>Comet Score (%)</td></tr><tr><td>English-Bengali</td><td>74.39</td></tr><tr><td>English-German</td><td>76.93</td></tr><tr><td>English-Greek</td><td>76.64</td></tr><tr><td>English-Hindi</td><td>67.87</td></tr><tr><td>English-Italian</td><td>79.04</td></tr><tr><td>English-Nepali</td><td>86.84</td></tr><tr><td>English-Russian</td><td>79.87</td></tr><tr><td>English-Swahili</td><td>73.71</td></tr><tr><td>English-Telugu</td><td>69.02</td></tr><tr><td>English-Bengali</td><td>78.79</td></tr><tr><td>English-Turkish</td><td>74.63</td></tr><tr><td>English-Chinese</td><td>74.63</td></tr><tr><td>English-Portuguese</td><td>74.89</td></tr><tr><td>System Score</td><td>76.05</td></tr></table>",
+        "bbox": [
+            144,
+            152,
+            455,
+            449
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Table 7: Average score from CometWiki of the Automatic Translation Quality without reference. The high resource languages (i.e., Italian, Greek etc) have higher scores than lower resource languages (i.e., Telugu)",
+        "bbox": [
+            112,
+            458,
+            489,
+            517
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "page_number",
+        "text": "4169",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "D Complete Results for English and Multilingual Experiments",
+        "text_level": 1,
+        "bbox": [
+            114,
+            83,
+            442,
+            118
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "We observed the mean accuracy of the MFC evaluation set for models trained on English and Multilingual datasets. The key findings are summarized below:",
+        "bbox": [
+            112,
+            124,
+            489,
+            187
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. The MFC alone achieved higher accuracy compared to other systems, with scores of  $61.93\\%$  and  $69.52\\%$  for BERT and RoBERTa-based models, respectively. However, when using the MFC10 dataset with limited high-quality data, the accuracy dropped significantly to  $53.02\\%$  and  $57.45\\%$  for BERT and RoBERTa models, respectively.",
+            "2. The SNFC and MaSNFC datasets exhibited lower accuracy when evaluated individually, compared to the MFC. However, the SNFC outperformed MFC10 in terms of accuracy for the BERT model. The SNFC has an accuracy of  $60.57\\%$  while the MFC10 has gotten  $53.02\\%$ . It is worth noting that the larger size of the SNFC contributed to its higher accuracy compared to MaSNFC, which is almost three times smaller.",
+            "3. Combining the MFC with our datasets led to substantial accuracy improvements. The models trained on MFC+SNFC (72.57%, 72.07%) and MFC+MaSNFC (72.85%, 73.22%) achieved higher accuracy than the MFC alone (61.93%, 69.52%), for both BERT and RoBERTa models.",
+            "4. Combining MFC10 with our datasets, we observed improved accuracy as well. The MFC10+SNFC combination yielded an accuracy improvement of 6.1 and 4.77 percentage points for BERT and RoBERTa models, respectively, compared to MFC10. Similarly, MFC10+MaSNFC demonstrated a similar improvement of 7.1 and 3.49 percentage points, respectively.",
+            "5. The overall accuracies of the MFC evaluation set for multilingual data (Table 3) are lower compared to the accuracies for English training (Table 2). This can be attributed to the fact that the training data in other languages were obtained through automatic translation, which may not be of the same quality as human translations or original news articles in those languages."
+        ],
+        "bbox": [
+            127,
+            195,
+            489,
+            921
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "table",
+        "img_path": "images/66357f14c2d42a340b50b56fd46ae623b05b07a27261371ba6608d2b5e0b4d20.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>System Name</td><td>Number of Sentences</td><td>BERT</td><td>RoBERTa</td></tr><tr><td>MFC</td><td>9740</td><td>61.93</td><td>69.52</td></tr><tr><td>MFC10</td><td>1125</td><td>53.02</td><td>57.45</td></tr><tr><td>SNFC</td><td>17520</td><td>60.57</td><td>54.37</td></tr><tr><td>MaSNFC</td><td>5182</td><td>52.05</td><td>48.77</td></tr><tr><td>MFC+</td><td>27260</td><td>72.57</td><td>72.07</td></tr><tr><td>SNFC</td><td></td><td></td><td></td></tr><tr><td>MFC+</td><td>14922</td><td>72.85</td><td>73.22</td></tr><tr><td>MaSNFC</td><td></td><td></td><td></td></tr><tr><td>MFC10+</td><td>18645</td><td>68.03</td><td>64.75</td></tr><tr><td>SNFC</td><td></td><td></td><td></td></tr><tr><td>MFC10+</td><td>6307</td><td>60.12</td><td>60.94</td></tr><tr><td>MaSNFC</td><td></td><td></td><td></td></tr></table>",
+        "bbox": [
+            514,
+            80,
+            878,
+            311
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "Table 8: Mean Accuracy Scores on the MFC evaluation set for models trained on English Datasets. The best scores have been highlighted.",
+        "bbox": [
+            507,
+            319,
+            882,
+            363
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "6. Among the datasets, MFC+MaSNFC achieved the highest accuracy of 45.73 on the multilingual test set, outperforming both MFC and MFC10 datasets.",
+            "7. For the Bengali test set, the highest accuracy (32.02) was achieved by the MFC10+SNFC training dataset. As for the Portuguese test set, the highest accuracy of 33.61 was obtained by two systems: MFC and MFC+MaSNFC.",
+            "8. Overall, the accuracies for the Bengali and Portuguese test sets were lower than those for the MFC evaluation set. This can be attributed to two factors. First, the training data, being translations, may not capture the nuances of the original news articles. Second, the training data mainly consists of MFC, which is collected from US-based news media sources. The test sets, on the other hand, were collected from Brazil and Bangladesh, which have different cultural contexts in their news articles that cannot be fully replicated through translation. To improve the scores further, it would be necessary to obtain original news articles from diverse culturally distinct sources in different languages."
+        ],
+        "bbox": [
+            524,
+            388,
+            884,
+            813
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "The study highlights challenges in multilingual framing analysis, with lower accuracies compared to English training. It emphasizes the need for high-quality translations and original news articles. Combining datasets like MFC+MaSNFC can enhance accuracy. Considering cultural and linguistic con",
+        "bbox": [
+            507,
+            824,
+            882,
+            921
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "page_number",
+        "text": "4170",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "table",
+        "img_path": "images/fc511cd82b864c70df79850d8097914476f63f0f25a27cd49f46eb7dba494f49.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">System Name</td><td colspan=\"2\">MFC Evaluation Set</td><td colspan=\"2\">Bengali Test Set</td><td colspan=\"2\">Portuguese Test Set</td></tr><tr><td>mBERT</td><td>XLM-R</td><td>mBERT</td><td>XLM-R</td><td>mBERT</td><td>XLM-R</td></tr><tr><td>MFC (English)</td><td>27.70</td><td>28.13</td><td>16.67</td><td>25.44</td><td>26.23</td><td>28.28</td></tr><tr><td>MFC</td><td>44.87</td><td>44.99</td><td>21.93</td><td>25.88</td><td>30.33</td><td>33.61</td></tr><tr><td>MFC10</td><td>27.7</td><td>28.64</td><td>20.61</td><td>23.68</td><td>30.33</td><td>27.87</td></tr><tr><td>SNFC</td><td>28.05</td><td>28.04</td><td>22.37</td><td>25.44</td><td>27.05</td><td>23.77</td></tr><tr><td>MaSNFC</td><td>28.86</td><td>27.55</td><td>11.84</td><td>16.67</td><td>20.49</td><td>15.98</td></tr><tr><td>MFC+SNFC</td><td>45.09</td><td>44.07</td><td>23.25</td><td>26.31</td><td>29.92</td><td>31.56</td></tr><tr><td>MFC+MaSNFC</td><td>44.42</td><td>45.73</td><td>22.37</td><td>28.07</td><td>31.97</td><td>33.61</td></tr><tr><td>MFC10 + SNFC</td><td>30.01</td><td>33.11</td><td>25</td><td>32.02</td><td>29.51</td><td>26.62</td></tr><tr><td>MFC10+MaSNFC</td><td>33.33</td><td>32.56</td><td>22.81</td><td>24.56</td><td>22.13</td><td>26.64</td></tr></table>",
+        "bbox": [
+            139,
+            80,
+            858,
+            262
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "text",
+        "text": "Table 9: Mean Accuracy Scores on the MFC evaluation set and Novel Multilingual Test Set for models trained on Multilingual Datasets. The best scores have been highlighted.",
+        "bbox": [
+            112,
+            272,
+            882,
+            302
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "text",
+        "text": "texts and diverse training data is crucial for better understanding framing across languages and cultures.",
+        "bbox": [
+            112,
+            326,
+            489,
+            374
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "page_number",
+        "text": "4171",
+        "bbox": [
+            480,
+            928,
+            517,
+            940
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "text",
+        "text": "E Instruction for the Generative AI Models",
+        "text_level": 1,
+        "bbox": [
+            114,
+            83,
+            438,
+            116
+        ],
+        "page_idx": 16
+    },
+    {
+        "type": "text",
+        "text": "This was the instruction that was given to the models discussed in Section 5.",
+        "bbox": [
+            112,
+            131,
+            485,
+            162
+        ],
+        "page_idx": 16
+    },
+    {
+        "type": "text",
+        "text": "\"In this task, you will be provided with a list of frames and a sentence. Your goal is to select the single most suitable frame from the given list for the provided sentence. Frames are cognitive structures that help humans interpret information by providing a mental framework for understanding. Each frame represents a specific perspective, context, or interpretation. Frame Selection Format: In your response, do not write anything other than the name of the frame. Frames List and Definitions: 'Economic': 'The financial consequences and economic implications of the matter on various levels (person, family, community or broader economy)'. 'External Regulation and Reputation': 'The external relations of nations or groups, trade agreements, policy outcomes, and external perceptions or consequences.' 'Political': 'Political considerations, actions, efforts, stances, and partisan, bipartisan, or lobbying activities related to an issue.' 'Public Sentiment': 'References of attitudes and opinions of the general public, including polling and demographics.' 'Cultural Identity': 'Traditions, customs, or values of a social group in relation to a policy issue.' 'Quality of Life': 'Threats and opportunities for the individual's wealth, happiness, and well-being.' 'Health and Safety': 'Access to healthcare, health outcomes, disease, sanitation, mental health, violence prevention, infrastructure safety, and public health.' 'Security and Defense': 'Actions or calls to action aimed at protecting individuals, groups, or nations from potential threats to their well-being.' 'Crime and Punishment': 'Effectiveness and implications of laws and their enforcement.' 'Policy Prescription and Evaluation': 'Specific policies proposed to address identified issues and the assessment of policy effectiveness.' 'Legality, Constitutionality, Jurisdiction': 'Discusses rights, freedoms, and authority of individuals, corporations, and government.' 'Fairness and Equality': 'The balance or distribution of laws, rights, and resources among individuals or groups.' 'Morality': 'Perspectives, policy objectives, or actions driven by religious principles, duties, ethics, or social responsibilities.' 'Capacity and Resources': 'The presence or",
+        "bbox": [
+            112,
+            179,
+            490,
+            921
+        ],
+        "page_idx": 16
+    },
+    {
+        "type": "text",
+        "text": "absence of various resources (physical, geographic, human, and financial) and the ability of existing systems.,'Other': 'Frames that don't fit into the categories above.' Please select the most appropriate frame for the given sentence, and specify the chosen frame without additional commentary. You are allowed to choose only one frame from the provided list that best aligns with the meaning and context of the given sentence. Consider the nuances of the sentence and the various frames in the list before making your selection. Your choice should reflect the frame that you believe most accurately captures the intended interpretation of the sentence. Remember that frames can significantly impact how information is perceived and understood. Choose the frame that enhances the sentence's meaning and effectively shapes the interpretation. Your ability to accurately match sentences with appropriate frames will contribute to successful communication and understanding. Do not write anything like, 'The selected frame is.' Input: \"",
+        "bbox": [
+            507,
+            84,
+            885,
+            439
+        ],
+        "page_idx": 16
+    },
+    {
+        "type": "text",
+        "text": "F Erroneous Predictions from Generative LLMs",
+        "text_level": 1,
+        "bbox": [
+            509,
+            450,
+            882,
+            482
+        ],
+        "page_idx": 16
+    },
+    {
+        "type": "page_number",
+        "text": "4172",
+        "bbox": [
+            480,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 16
+    },
+    {
+        "type": "table",
+        "img_path": "images/8ca37ec36d6879e47d39156bfcee6a8591849e3768ed002bf1cad56cdbed7813.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Text</td><td>Gold Label Output</td><td>Models</td><td>Model Prediction</td></tr><tr><td>DETENTION OF REFUGEEES IN TEXAS OK&#x27;D</td><td>Crime and Punishment</td><td>Llama 70B</td><td>The External Regulations And reputation was picked since it refers nation borders control migrants entering its territory as refugee status;</td></tr><tr><td>But last spring, the attorney for the regents, Burns Newsome, advised them to end the waivers for illegal immigrants, saying the practice is at odds with SB 529. Now, the regents are having a series of public hearings to get input on the issue.</td><td>Legality, Constitu-tionality, Jurisdiction</td><td>Llama 70B</td><td>External Regulation and Renown</td></tr><tr><td>3 Prison Guards Guilty of Abuse Of Immigrants</td><td>Crime and Punishment</td><td>Falcon 40B</td><td>Prison guards guilty.</td></tr><tr><td>DHS budget brinkmanship could leave local governments in the lurch</td><td>Economic</td><td>Falcon 40B</td><td>Economical</td></tr><tr><td>Disorder On The Border</td><td>Security and Defense</td><td>Mistral 7B</td><td>The most suitable frame for the sentence ‘Disorder on the Border’ is ‘Security and Defense’</td></tr><tr><td>Police say Orleans-Lindsay shot Toatley, 37, to death Oct. 30 during an undercover drug purchase. Orleans-Lindsay, of Silver Spring, allegedly accepted $3,500 in cash for cocaine, pretended to get the drugs and returned to the trooper&#x27;s sport-utility vehicle a few minutes later, firing a single shot. Toatley, a 16-year trooper and father of three, died 2 1/2 hours later.</td><td>Crime and Punishment</td><td>Mistral 7B</td><td>‘Law Enforcement’, ‘Political’, ‘Fair-ness and Equality’, ‘Crime and Punishment’ and ‘Legality, Constitutionality, Jurisdiction’</td></tr></table>",
+        "bbox": [
+            122,
+            179,
+            875,
+            778
+        ],
+        "page_idx": 17
+    },
+    {
+        "type": "text",
+        "text": "Table 10: LLMs predictions compared to gold label outputs. Errors include spelling errors to not following instruction.",
+        "bbox": [
+            112,
+            788,
+            882,
+            816
+        ],
+        "page_idx": 17
+    },
+    {
+        "type": "page_number",
+        "text": "4173",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 17
+    }
+]

2024/A Study on Scaling Up Multilingual News Framing Analysis/a4b2e349-9396-4656-a45a-c4d7d00460a3_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:31990842c1044da4faf0e3563c346f701167246258d2e92e667e3f14ff4a9e28
+size 739222

2024/A Study on Scaling Up Multilingual News Framing Analysis/full.md

@@ -0,0 +1,309 @@
+# A Study on Scaling Up Multilingual News Framing Analysis
+
+Syeda Sabrina Akter and Antonios Anastasopoulos  
+Department of Computer Science, George Mason University  
+{sakter6, antonis}@gmu.edu
+
+# Abstract
+
+Media framing is the study of strategically selecting and presenting specific aspects of political issues to shape public opinion. Despite its relevance to almost all societies around the world, research has been limited due to the lack of available datasets and other resources. This study explores the possibility of dataset creation through crowdsourcing, utilizing non-expert annotators to develop training corpora. We first extend framing analysis beyond English news to a multilingual context (12 typologically diverse languages) through automatic translation. We also present a novel benchmark in Bengali and Portuguese on the immigration and same-sex marriage domains. Additionally, we show that a system trained on our crowdsourced dataset, combined with other existing ones, leads to a 5.32 percentage point increase from the baseline, showing that crowdsourcing is a viable option. Last, we study the performance of large language models (LLMs) for this task, finding that task-specific fine-tuning is a better approach than employing bigger non-specialized models. $^{1}$
+
+# 1 Introduction
+
+News framing refers to the power of the news media to define and interpret events, issues, and policies by emphasizing certain aspects while downplaying or excluding others. According to Entman (1993), it can "make a piece of information more noticeable, meaningful, or memorable to audiences". It plays a crucial role in influencing how people interpret and react to information presented in news articles. The language used in news media can shape public opinion and reveal biases and agendas, which can ultimately shape the way people understand and react to current events.
+
+- A presidente do País, Tsai Ing-wen, également se pronunciou a favor da lei e pediu acos legisladores empenho com a pauta Political. "Temos a opportunidade de fazer-history e做不到 ao mundo que valeores progressistas podemcriar raízes nas sociedades da Ásia Oriental", affirmou Fairness and Equality.
+
+Figure 1: The image illustrates the process of framing in Portuguese at the sentence level, showcasing how specific language for each sentence strategically shape a Political and Equality narrative in the same article.
+
+Traditionally, framing analysis has relied on manual annotation by linguists, social studies experts, and trained annotators, lacking the potential of AI-driven systems leading to a rather limited explorations of automating framing analysis. Moreover, existing studies have been restricted primarily to English-only data, leaving a gap in research concerning multilingual and low-resource contexts.
+
+Our work focuses on employing NLP techniques for the framing analysis task to automate the analysis process, extract insights from large datasets efficiently, and identify patterns in the language used in news media. To address these challenges, Boydstun et al. (2014) introduced a codebook, Policy Frames Codebook, based on which the Media Frames Corpus (MFC; Card et al., 2015) was created. This dataset is comprised broad categories of common policy frames and annotations of US news articles. However, the availability of such datasets in languages beyond English remains limited.
+
+Getting a higher volume of higher quality data (such as, MFC) is time and resource intensive. Hence, we study the alternative of gathering a high volume of comparatively lower quality but easy-to-collect data. We achieve this through crowdsourcing and automatic translation techniques. We also examine the combination of lower and higher quality data.
+
+In this study, we first introduce a new crowdsourced dataset: Student-sourced Noisy Frames
+
+Corpus (SNFC). We have achieved time and cost efficiency by involving a large number of semitrained annotators for the data collection and annotation process of the corpus. SNFC covers immigration and same-sex marriage domains and includes novel benchmark test sets in Bengali and Portuguese, offering new perspectives in these languages. Additionally, we automatically expand multilinguality to the task by translating the MFC and SNFC to 12 more languages. We show that a neural classifier trained on the combination of both MFC and SNFC yields significant performance improvements, both in English as well as in a multilingual setting. Finally, we explore generative large language models, such as LLaMA (Touvron et al., 2023), to study their efficacy for this task.
+
+Our findings show that neural models trained on SNFC can reach the performance levels of those trained on high quality data (i.e., MFC). Going further, we find that the combination of expert and non-expert annotated data (i.e. MaSNFC+MFC) outperforms just MFC, which provides a path towards expanding coverage without the need for expensive expert annotations.
+
+# 2 Related Work
+
+Framing analysis provides valuable insights into different perspectives on news topics across various countries and languages. However, there is a notable lack of research and annotated corpora for framing analysis in languages other than English. This limitation hinders our understanding of media framing in different parts of the world and other societies' opinion regarding specific issues. To address this gap, a multilingual approach is essential in analyzing media framing across diverse linguistic and cultural contexts. Ali and Hassan (2022) provide a comprehensive survey of the framing analysis task, focusing specifically on studies in English datasets exploring various approaches and techniques employed in framing analysis.
+
+Two prominent datasets used for framing analysis are the Media Frames Corpus (MFC; Card et al., 2015) and the Gun Violence Frames Corpus (GVFC; Liu et al., 2019). The MFC, annotated according to the guidelines provided in the codebook of Boydstun et al. (2014), covers 6 different political issues including immigration, same-sex marriage, and gun violence, among others. It includes both article headlines and news texts, providing a broader and more comprehensive dataset.
+
+On the other hand, the GVFC focuses solely on the topic of gun violence, with 10 manually annotated frames defined in a different codebook, and it only includes article headlines.
+
+Akyurek et al. (2020) extended the GVFC by curating headlines in German, Turkish, and Arabic following the same process as the original dataset from the respective news websites, specifically targeting keywords related to gun violence and mass shootings. The frames used in the multilingual datasets remained consistent with those in the GVFC, and is the one of the few multilingual sources for this task. Additionally, the Australian Parliamentary Speeches (APS) dataset (Khanehzar et al., 2019) offers another perspective on framing analysis, as it consists of transcripts speeches related to same-sex marriage bills presented in the Australian Parliament. Although the APS dataset focuses on data from a country other than the United States, it is still limited to English language texts, which narrows the scope of the framing analysis task.
+
+The MFC has served as a valuable resource in various framing-related studies. For example, it was used to develop a semi-supervised model by extracting a Russian lexicon from their Russian test corpora which consists of news articles sourced from reputable Russian newspapers (Field et al., 2018). In a different vein, Naderi and Hirst (2017) used it to benchmark sentence-level classification tasks, employing LSTM, BiLSTM, and GRU-based systems. Considering the significant contributions of this corpus to the field, we have incorporated it into our system for training and evaluation purposes, alongside our SNFC dataset.
+
+Several studies have employed various techniques such as topic modeling (DiMaggio et al., 2013; Roberts et al., 2014; Nguyen, 2015), cluster analysis (Burscher et al., 2016), and neural networks (Naderi and Hirst, 2017; Khanehzar et al., 2019; Mendelsohn et al., 2021; Kwak et al., 2020) to construct systems for framing analysis. These investigations have consistently demonstrated that leveraging state-of-the-art pre-trained models based on transformers (Devlin et al., 2019; Zhuang et al., 2021; Conneau et al., 2020) is a highly effective approach, yielding significantly improved results compared to other techniques. In our study, we follow the state of the art and build models similar to those employed by Liu et al. (2019) and Khanehzar et al. (2019).
+
+We also investigated crowdsourcing methods
+
+which, as defined by Howe (2006), is an online, distributed problem-solving and production model that leverages the collective intelligence of online communities for specific goals. This technique aims to tap into the global talent pool, accelerating innovation and problem-solving across various domains. Hossain and Kauranen (2015) provide a comprehensive literature review, identifying numerous crowdsourcing methods, which emphasizes the difficulty of generalizing these methods due to their diversity and application-specific nature. However, the widespread use of these methods demonstrates versatility and adaptability of different crowdsourcing methods. Zhao and Zhu (2014) suggest that future research should focus on standardizing crowdsourcing processes to enhance efficiency and effectiveness. This indicates an increasing realization of the necessity to codify crowdsourcing approaches, notwithstanding their inherent variability.
+
+# 3 Dataset Creation
+
+In this section, we present our methodology for curating SNFC training dataset through crowdsourcing (§3) and outline the process of extending the dataset to incorporate multilinguality (§3). Lastly, we introduce our innovative Portuguese and Bengali benchmarks, highlighting their significance in the context of this study (§3).
+
+SNFC Training Corpus To construct the crowdsourced training portion of the SNFC, we turned to students at George Mason University. In particular, this was done as part of an in-class assignment for a graduate-level natural language processing class with about 80 students involved.[2]
+
+The students were presented with the challenge of building a Media Frames Analysis system (effectively, a sentence-level neural classifier), without having access to significant amounts of data. In particular, the students were provided only with a description of the codebook of Boydstun et al. (2014) presented in Table 5, along with 250 sentence-level examples called the seed dataset from the MFC corpus sampled so that all 15 frame dimensions were present.
+
+The codebook and the samples were meant to facilitate the annotators' understanding of the task. The only other information available to them was that their final systems would be evaluated on multiple languages (see §3) on the immigration and
+
+same-sex marriage domains.3
+
+The students were first tasked with procuring 150 new sentences each, from any source and in any language, and label them, according to the codebook, to be used as their "first" training set. They then had to produce an additional 150 sentences which would then be annotated by two of their peers (so that we will be able to measure inter-annotator agreement). Any label disagreements were resolved by the students, by obtaining an additional label for majority voting. All in all, each student produced a minimum of 300 annotated sentences. While the students had the option to collect data in any language, all of them, apart from two, collected and annotated the initial data in English. The two other students who collected data in different languages chose their native languages: Telugu, and Hindi.
+
+To collect the data, the students were allowed to do anything they wanted. They ended up utilizing diverse techniques that range from targeted web scraping to generating sentences with the assistance of AI tools such as, ChatGPT (Radford et al., 2019). We can broadly categorize the sources of data into three categories: AI tools (such as ChatGPT and ChatSonic), online news platforms (including Online Articles, NBC, CNN, BBC, and NYTimes), and social media platforms (such as Twitter and Reddit). Students have used a combination of two or more categories to collect their data. Around  $77\%$  of students used AI tools,  $14.8\%$  relied on social media platforms, and  $67.9\%$  used online news platforms for data collection purposes. It is important to note that, AI was only used by the students in the first step of data collection. This shows how artificial intelligence (AI) eases the process of collecting relevant, topic-specific text. The process of data validation and labeling was entirely done by human annotators.
+
+In the end, we ended up with a total of 17,520 sentences from the combined student training corpus of 300 sentences each, eliminating the occasional duplicate instances. The dataset has a generally substantial inter-annotator agreement, with a Cohen's  $\kappa$  (Cohen, 1960) coefficient of 0.61.
+
+To further contextualize this, we note that the inter-annotator agreement of the MFC (as detailed in the paper) is assessed using Krippendorff's  $\alpha$  (Krippendorff, 2011), with respective values of 0.08 and 0.20 for the domains of same-sex marriage and im
+
+migration. SNFC (our dataset) combines sentences from both of these domains and the Krippendorff's  $\alpha$  value for SNFC stands at 0.103 which is similar to the one of MFC. Given that this is a 15-way classification task, we believe the inter-annotator agreement for SNFC is not particularly low for such a nuanced task.
+
+Multilinguality To benchmark media framing beyond English our first step is to simply translate the original MFC dataset into other languages. We use machine translation<sup>4</sup> to translate all sentences of the MFC corpus into 12 typologically diverse languages, namely Bengali, German, Greek, Italian, Turkish, Nepali, Hindi, Portuguese, Telugu, Russian, Swahili, and Mandarin Chinese.
+
+While the primary reason for this process is the ability to benchmark the task on other languages (as well as the inability to collect annotated test sets in all of these languages - see also §3), this simple data augmentation technique is also a reasonable way to also obtain training data in other languages. Hence, we perform this translation both on the training and the dev/test portions of the dataset, and combine all languages to form the multilingual version of the dataset.
+
+Lastly, the same translation models were used to augment our crowd-sourced SNFC dataset to cover all of the above-mentioned languages.
+
+We have studied the quality of the translation through human assessment. For each language, we took 100 translations from English and had them reviewed by bilingual speakers who scored the translations on a scale from 1 to 10 based on accuracy and clarity. For this evaluation, we used four languages: Bengali, Greek, Hindi, and Nepali. From the average rating for each language pair (See Table 1), we observe that the average rating is higher for higher resourced languages like Greek and Hindi. On the other hand, Nepali, being the only lower resourced language, has a lower rating of 4.72 out of 10, suggesting that perhaps Nepali results should be taken with a grain of salt, as the reason for general poor performance is likely to be the low quality of the translations.
+
+We have also further performed quality estimation over all translations by calculating the CometKiwi score (Rei et al., 2023) of the translations. Note that we resort to automatic quality estimation since we do not have access to reference translations. The overall score of  $76.05\%$  is
+
+<table><tr><td>Language Pair</td><td>Rating (%)</td></tr><tr><td>English-Bengali</td><td>61.2</td></tr><tr><td>English-Greek</td><td>73.4</td></tr><tr><td>English-Hindi</td><td>77.4</td></tr><tr><td>English-Nepali</td><td>47.2</td></tr><tr><td>Comet Score (All languages)</td><td>76.05</td></tr></table>
+
+Table 1: Average rating for Human Evaluation of the Automatic Translation Quality
+
+in line with our human evaluation over the sample, and suggests that automatic translations are largely reliable in our dataset. The higher scores for the high resource languages of the human-evaluation and CometKiwi (see Appendix C for a breakdown by language) indicate that automatic translations can be a reasonable alternative to gathering large quantities of high quality multilingual data for the framing task.
+
+Novel Test Set While the automatic translation of the MFC benchmark is a reasonable start for our multilingual exploration, it does not come without drawbacks: the provided text, regardless of the language, is only relevant to the USA cultural context.
+
+To even better benchmark the quality of framing analysis systems on different language and cultural contexts, we create a pair of novel test sets in (Bangladesh) Bengali and (Brazilian) Portuguese. The news articles used in this test set were sourced from reputable newspapers in Bangladesh and Brazil, aligning with the chosen domains of immigration and same-sex marriage.
+
+Each test set is comprised of 10 news articles for each language. The annotators were native speakers of the languages and they adhered closely to the definitions provided by the authors (Table 5), ensuring consistency with the labels found in the MFC.
+
+Figure 2 shows the label distribution for the MFC and the novel test set, listing the number of sentences per frame in each language. In the case of Bengali, the news articles predominantly focus on the immigration domain, reflecting the cultural disparities between Brazil and Bangladesh. Specifically, the test set emphasizes the economic and lifestyle aspects of immigration (Bengali), while also delving into the legal and policy-making dimensions of the domain (Portuguese).
+
+It is of note that the two benchmarks, despite being rather small, still show interesting differences in terms of their label distribution. For example, the
+
+![](images/bb5830acf1a76884af095a52ca26b276aaeb48bb2acfa5bcee9071befdb2be4e.jpg)  
+Figure 2: The label distributions of the MFC and our new Bengali and Portuguese test sets. Note that they differ significantly.
+
+most common label on the Bengali set is "External Regulation and Reputation", which is the least common one in the Portuguese one. And the reverse is the case for the "Cultural Identity" label which is the most common in Portuguese and least common in Bengali. Another interesting observation is that the Bengali test set contains more data labeled as "Other" compared to the other two languages. Upon analyzing the data with the help of a native speaker, we found that most of the Bangladeshi articles emphasize a lot on reporting information in the form of dates and numbers, rather than offering opinions on the issues.
+
+# 4 Framing Analysis System and Results
+
+Experimental Setup We approach the task as a multilabel classification problem (Tsoumakas and Katakis, 2007), leveraging the pretrained RoBERTa (Zhuang et al., 2021) language model, similar to the SOTA approach employed by Khanehzar et al. (2019). For all models we set the maximum sequence length to 256, with a batch size of 16, and train using a learning rate of  $10^{-5}$ . To expand to more languages, we employ the multilingual XLM-RoBERTa model (Conneau et al., 2020). Throughout all experiments, we use the base model size.
+
+We first report results with models exclusively trained on MFC, and SNFC datasets, as well as their concatenation. To investigate a more data-scarce scenario, we also compiled a smaller sample consisting of about  $10\%$  of the original MFC,
+
+<table><tr><td>Tr. Data</td><td>#Sentences</td><td>Accuracy</td></tr><tr><td colspan="3">Baselines</td></tr><tr><td>MFC</td><td>9739</td><td>69.52</td></tr><tr><td>MFC10</td><td>1125</td><td>57.45</td></tr><tr><td colspan="3">including crowd-sourced data</td></tr><tr><td>SNFC</td><td>17520</td><td>54.37</td></tr><tr><td>SNFC50</td><td>8760</td><td>54.7</td></tr><tr><td>MFC+SNFC</td><td>27260</td><td>72.07</td></tr><tr><td>MFC+SNFC50</td><td>18499</td><td>72.89</td></tr><tr><td>MFC10+SNFC</td><td>18645</td><td>64.75</td></tr><tr><td>MFC10+SNFC50</td><td>9885</td><td>62.05</td></tr><tr><td colspan="3">filtered crowd-sourced data</td></tr><tr><td>MaSNFC</td><td>5182</td><td>48.77</td></tr><tr><td>MFC+MaSNFC</td><td>14922</td><td>73.22</td></tr><tr><td>MFC10+MaSNFC</td><td>6307</td><td>60.94</td></tr></table>
+
+Table 2: Mean Accuracy Scores on the MFC evaluation set for RoBERTa models trained on English Datasets. # stands for "number of".
+
+named MFC10, ensuring all 15 target labels are included. Beyond the single-dataset baselines, we combine the expert-annotated MFC and MFC10 with our crowd-sourced SNFC. To further study the effect of the size of the SNFC, we have experimented with SNFC50, a randomized halved subset of the original SNFC that is more closer to the MFC in size.
+
+English Results and Discussion We first establish the usefulness of our crowdsourced data, by focusing on the performance on the original test set of the English MFC dataset (using the monolingual RoBERTa model). Results are presented in Table 2.
+
+First, it is worth pointing out that relying solely on crowd-sourced data is not promising: the SNFC-only training underperforms both the MFC-only set
+
+ting, as well as the MFC10-only setting, which has only around  $10\%$  of the training data size!
+
+However, combining the expert-annotated data with the crowd-sourced ones yields significant improvements over the expert-only baselines, as MFC+SNFC yields an extra 2.5 accuracy points over MFC (72% vs 69.5%). The improvement is even larger (more than 7 accuracy points) in the resource-restricted MFC10 scenario. The accuracy remains consistent both with SNFC50 alone and when combined with MFC, as MFC+SNFC50 and MFC+SNFC yield similar results, indicating that performance gains are not merely due to larger data volume.
+
+Filtering of Crowdsourced Data Given the potential for noise in any crowd-sourced dataset, we explore a simple filtering technique to sample more high-quality crowd-sourced. In particular, we obtain sentence-level representations for each sentence, and select only the SNFC instances that exhibit more than  $85\%$  cosine similarity with any MFC instance. Effectively, we select SNFC sentences that are most similar to MFC ones. We refer to this sample as MFC-aligned SNFC (MaSNFC).
+
+Results with this (almost 3x smaller) sample are more encouraging (Table 2): combining MaSNFC with MFC yields our best model with an accuracy of 73.22. In the data-scarce scenario of MFC10, adding MaSNFC is again beneficial, but including the whole unfiltered SNFC is even better.
+
+These findings underline the promise of crowdsourcing for collecting a high volume of (somewhat) lower quality data. The performance improvement for the MaSNFC+MFC shows promise for the combination of low-volume high-quality along with a higher-volume of lower-quality data. This approach effectively balances the depth and breadth of the dataset, leveraging the strengths of both data types.
+
+Multilingual Results and Discussion For the first part of our multilingual experiments, we employ a translate-train and translate-test scenario. All of the dataset samples introduced above were translated to all 12 evaluation languages, and we now replicate the same experimental setups as above, the only difference being that we will use a multilingual LM (XLM-R instead of RoBERTa). All results are presented in Table 3 (which presents the average accuracy across the 12 languages for mMFC, as well as performance on our novel Bengali and Portuguese benchmark).
+
+<table><tr><td>Tr. Data</td><td>mMFC</td><td>BENGALI</td><td>PORTUGUESE</td></tr><tr><td colspan="4">Zero-shot (only English train)</td></tr><tr><td>MFC</td><td>28.13</td><td>25.44</td><td>28.28</td></tr><tr><td colspan="4">Baselines (translate-train)</td></tr><tr><td>MFC</td><td>44.99</td><td>25.88</td><td>33.61</td></tr><tr><td>MFC10</td><td>28.64</td><td>23.68</td><td>27.87</td></tr><tr><td colspan="4">+ crowd-sourced (translate-train)</td></tr><tr><td>SNFC</td><td>28.04</td><td>25.44</td><td>23.77</td></tr><tr><td>MFC+SNFC</td><td>44.07</td><td>26.31</td><td>31.56</td></tr><tr><td>MFC10+SNFC</td><td>33.11</td><td>32.02</td><td>26.62</td></tr><tr><td colspan="4">+ filtered crowd-sourced (translate-train)</td></tr><tr><td>MaSNFC</td><td>27.55</td><td>16.67</td><td>15.98</td></tr><tr><td>MFC+MaSNFC</td><td>45.73</td><td>28.07</td><td>33.61</td></tr><tr><td>MFC10+MaSNFC</td><td>32.56</td><td>24.56</td><td>26.64</td></tr></table>
+
+Table 3: Mean Accuracy Scores on the MFC evaluation set and Novel Multilingual Test Set for XLM-R models trained on Multilingual Datasets. The best scores have been highlighted.
+
+First of all, we show that relying on zero-shot cross-lingual transfer, without employing the translate-train technique is not a competitive baseline. The translated MFC baseline is competitive on average, but as we discuss below it performs quite inequitably across languages. As before, combining expert annotated data with filtered crowdsourced ones (MFC+MaSNFC) is best. Our findings from the monolingual experiments generally hold in the multilingual ones.
+
+In the Bengali test set, the inclusion of all crowdsourced data improves upon the baseline by a small margin. The improvement from filtered crowdsourced data is more modest. However, it is interesting that the best performance is obtained when using fewer expert annotations (MFC10+SNFC), improving by almost 6 percentage points over the baseline! We hypothesize that using the whole MFC dataset overfits the US context – but we leave this analysis for future work. In the Portuguese test set, we observe generally similar patterns as in the mMFC, with the exception that we do not observe any improvement from the crowd-sourced data. We leave a further investigation for future work.
+
+We note that the accuracies for the Bengali and Portuguese test sets are significantly lower than those of the English MFC and the mMFC test sets. We suspect that the training data, being automatic translations, may not capture the nuances of the original news articles. Second, the domain shift due to cultural context differences between training and test
+
+![](images/89f04374dd139fe0c5bae757792230560d2a822da3a6392020b9d15c164d7e49.jpg)  
+Figure 3: The best model performs very inequitably across languages on mMFC. The highest accuracy is in English (72.1%) followed by Italian and German, while other languages from non-western countries (e.g. Bengali, Hindi, Chinese, and others) have much lower performance (under 30%).
+
+may play a significant role. To improve the scores further, it may be necessary to obtain original news articles from diverse culturally distinct sources in different languages.
+
+mMFC Breakdown per Language We further analyse the per-language performance of our best-performing model on mMFC (see Figure 3). English accuracy (72.1) is en par with the monolingual setting (73.2), and German, Italian, Swedish, and Turkish also yield accuracies higher than  $64\%$ . But for other languages the model performs much worse, including high-resource ones like Greek  $(31.5\%)$ , Russian  $(28\%)$ , and Chinese  $(25.5\%)$ . While translation errors may play a role here, we are confident that they are not enough to explain such a large discrepancy. For example, while Nepali has admittedly low-quality translations (see previous discussion), Hindi, Greek, and Chinese certainly have translations of fairly high quality and yet they fall in the same low performance ballpark. We suspect that this gap may only be bridged through data collection (either expert- or crowd-annotated) in the appropriate languages and cultural contexts.
+
+Error Analysis We analyzed the errors using a confusion matrix for our best-performing model MFC+MaSNFC on the mMFC evaluation set, as shown in Figure 4. The heat-map reveals that out of 15 labels, 9 achieve the majority of instances correctly. Specifically, the labels 'Political' and 'Legality, Constitutionality, Jurisdiction' have the highest number of instances predicted correctly. However, when the model makes incorrect predictions, the errors are mainly categorized into the 'Political'
+
+and 'Legality, Constitutionality, Jurisdiction' labels. This led us to suspect a potential data imbalance in our training model. Further examination of the data confirmed that these two labels indeed have a majority of instances in the training set, leading to the tendency to predict these labels when uncertain.
+
+One could also further argue that these two labels are quite close semantically and hence their confusion is perhaps expected. We have examined the original data from MFC for the immigration and same-sex issues, which were used to train our baseline model. This dataset indeed shows a skewed distribution with a disproportionate number of instances falling under these two labels. This suggests that US-based news articles covering these domains inherently tend to fall in these two categories. Given the domain, we deduce that such an imbalance in label distribution might be a common trend in news articles from other countries as well. This assumption can be further validated in our novel test sets derived from Bangladesh and Brazil, which also reveal a similar inclination towards certain labels, as discussed in the previous section.
+
+# 5 Generative Language Models
+
+LLMs like GPT-4 (OpenAI, 2023), Falcon (Penedo et al., 2023), and LLaMA (Touvron et al., 2023), are trained on vast amounts of text and have shown immense promise in a variety of NLP tasks. Their broad knowledge base qualifies them as potential tools for framing analysis. In this study, we have also explored three of these models, particularly the open-sourced ones: Mistral, LLaMA-2, and Falcon.
+
+Figure 4: Confusion matrix for the best model's prediction for the mMFC Test set.  
+![](images/51f87766689f38ea67ee7eda04b48002e2204b35fdcfcb3a104deea857e9c7cc.jpg)  
+Economic: E  
+Capacity and Resources: CaR  
+Morality: M  
+Fairness and Equality: FaE  
+Legality, Constitutionality, Jurisdiction: LCJ  
+Policy Prescription and Evaluation: PPaE  
+Crime and Punishment: CaP  
+Security and Defense: SaD  
+Health and Safety: HaS  
+Quality of Life: QoL  
+Cultural Identity: CI  
+Public Sentiment: PS  
+Political: P  
+External Regulation and Reputation: ERaR  
+Other: O
+
+<table><tr><td>Model</td><td>Accuracy (%)</td></tr><tr><td>Falcon-40b-instruct</td><td>22.95</td></tr><tr><td>Mistral-7B-Instruct-v0.1</td><td>35.33</td></tr><tr><td>Llama2-chat-70B</td><td>22.22</td></tr></table>
+
+Table 4: Exact Match accuracy of the LLMs. The highest accuracy (35%, bolded) is significantly worse than the task finetuned RoBERTa model's performance (73.22%).
+
+Experimental Setting The instruction presents the framing task as a multiple choice question with 15 options and we have curated the instruction to include the definitions of all the labels, similar to the ones the students have used to annotate the SNFC. The instruction we use is given in Appendix E. We conduct all experiments in the zero shot setting, to assess the potential of LLMs to generalize and apply their knowledge effectively without task-specific training. The experiments were run on the English only test set (MFC-test) to ensure comparability with other task-finetuned models previously evaluated on the same test set.
+
+Results and Discussion The results (see Table 4) show the exact match accuracy of different LLMs on the MFC-test dataset. The performance of Llama2-chat-70B aligns closely with that of Falcon-40b-instruct, and Mistral-7B-Instruct-v0.1 outperformed them significantly showing that the sheer size of a model does not necessarily equate
+
+to better performance.
+
+Interestingly, the best performance was achieved by employing smaller, task-finetuned models, with RoBERTa achieving an exact match accuracy of  $73.22\%$ . This significantly surpasses the highest result for general LLMs, as their best performance is at  $35.33\%$ , observed with Mistral-7B-Instructv0.1. This difference in performance highlights the importance of task-specific fine-tuning on model efficacy. The finetuning process allows models like RoBERTa to adapt their parameters more closely to the nuances of the specific task, resulting in a more precise understanding and response generation compared to models that rely solely on broad, generalized training. The results also suggests that there is a trade-off between model size and specialized training. While larger models have a vast knowledge base, they are not always effective in applying this knowledge to specific tasks without fine-tuning.
+
+Error Analysis The LLMs exhibit a range of errors in predicting the correct frames for the provided texts (See Table 10). These errors include spelling mistakes, overgeneralization, assigning multiple labels where only one is appropriate, and misinterpretation. Generally, the models struggle with adhering to instructions, such as inventing new frames rather than selecting from the provided list (External Regulatory and Renown). Additionally, a common issue among all three of the models is
+
+their failure to introduce their answers concisely as instructed. Contrary to the clear direction to reply only with the label name, they begin responses with phrases like 'The most suitable frame is...'.
+
+The Mistral 7B model achieves a higher accuracy rate compared to the other two model; however, it often adds additional commentary to its responses. The LLaMA-2 70B model's predictions are inconsistent, notably when it replaces 'External Regulation and Reputation' with 'External Regulatory and Renown', demonstrating a tendency towards misrepresentation. The Falcon 40B sometimes accurately identifies the frame but fails to use the exact label name, responding with 'Economical' instead of 'Economic'.
+
+Since the models have the tendency to predict labels with spelling errors and synonymous labels, we have employed different techniques to measure the accuracy of these models to ensure a true reflection of the system's performance. To derive the correct label names from synonymous words and to overlook spelling mistakes, we employed the Fast-Text (Joulin et al., 2016) and Edit Distance (Levenshtein et al., 1966) algorithms. These were used to determine the textual similarity between the models' predictions and the 15 labels they were intended to predict.
+
+# 6 Conclusion
+
+In conclusion, our study emphasizes the importance of data quality and language diversity in multilingual framing analysis. Combining the Media Frames Corpus (MFC) with the Student-Sourced Noisy Frames Corpus (SNFC) yields significant improvements, highlighting the value of larger datasets despite the annotation quality potentially being lower. However, lower accuracy in multilingual experiments indicates the need for accurate translations and culturally diverse training data to improve multilingual framing analysis. Last, the sub-par performance of LLMs showcases a future research direction towards task-specific finetuning of the LLMs.
+
+# Limitations
+
+The main limitation of this study is that it relies on automated translation via Google Translator to introduce multilinguality to the task. It is well known that the translations conducted by Google Translator may not achieve the same level of quality as authentic translations. Moreover, for lower
+
+resource languages such as Nepali and Swahili, the translations obtained from Google Translator may not fully capture the nuances and characteristics as well as it probably can if translated to higher-resource languages as German or Greek. Additionally, since the MFC dataset primarily consists of US news sources, the translations into different languages does not adequately reflect the biases and perspectives surrounding a specific political issue in different countries. We attempt to mitigate this limitation with our new Bengali and Portuguese test sets. Collecting more data from different countries in different languages will eventually address this limitation, but we leave this large-scale undertaking for the future.
+
+# Acknowledgements
+
+We are thankful to the anonymous reviewers for their useful feedback, as well as to the students of the GMU CS 678 course and the annotators for the Bengali and Portuguese test set who majorly contributed in the creation of our crowdsourced dataset. This project was supported by the National Science Foundation under grant IIS-2327143. This project was also supported by resources provided by the Office of Research Computing at George Mason University (https://orc.gmu.edu) and funded in part by grants from the National Science Foundation (Awards Number 1625039 and 2018631).
+
+# References
+
+Afra Feyza Akyurek, Lei Guo, Randa I. Elanwar, Prakash Ishwar, Margrit Betke, and Derry Tanti Wijaya. 2020. Multi-label and multilingual news framing analysis. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 8614-8624. Association for Computational Linguistics.  
+Mohammad Ali and Naeemul Hassan. 2022. A survey of computational framing analysis approaches. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, EMNLP 2022, Abu Dhabi, United Arab Emirates, December 7-11, 2022, pages 9335-9348. Association for Computational Linguistics.  
+Amber E Boydstun, Dallas Card, Justin Gross, Paul Resnick, and Noah A Smith. 2014. Tracking the development of media frames within and across policy issues.  
+Bjorn Burscher, Rens Vliegenthart, and Claes H de Vreese. 2016. Frames beyond words: Applying cluster and sentiment analysis to news coverage of the
+
+nuclear power issue. Social Science Computer Review, 34(5):530-545.  
+Dallas Card, Amber E. Boydstun, Justin H. Gross, Philip Resnik, and Noah A. Smith. 2015. The media frames corpus: Annotations of frames across issues. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing of the Asian Federation of Natural Language Processing, ACL 2015, July 26-31, 2015, Beijing, China, Volume 2: Short Papers, pages 438-444. The Association for Computer Linguistics.  
+Jacob Cohen. 1960. A coefficient of agreement for nominal scales. Educational and psychological measurement, 20(1):37-46.  
+Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettle-moyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 8440-8451. Association for Computational Linguistics.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 4171-4186. Association for Computational Linguistics.  
+Paul DiMaggio, Manish Nag, and David Blei. 2013. Exploiting affinities between topic modeling and the sociological perspective on culture: Application to newspaper coverage of us government arts funding. Poetics, 41(6):570-606.  
+Robert M Entman. 1993. Framing: Toward clarification of a fractured paradigm. Journal of communication, 43(4):51-58.  
+Anjalie Field, Doron Kliger, Shuly Wintner, Jennifer Pan, Dan Jurafsky, and Yulia Tsvetkov. 2018. Framing and agenda-setting in russian news: a computational analysis of intricate political strategies. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Brussels, Belgium, October 31 - November 4, 2018, pages 3570-3580. Association for Computational Linguistics.  
+Mokter Hossain and Ilkka Kauranen. 2015. Crowdsourcing: a comprehensive literature review. Strategic Outsourcing: An International Journal, 8(1):2-22.  
+Jeff Howe. 2006. The rise of crowdsourcing, wired. http://www.wired.com/wired/archive/14.06/crowds.html.
+
+Armand Joulin, Edouard Grave, Piotr Bojanowski, Matthijs Douze, Hérve Jégou, and Tomas Mikolov. 2016. Fasttext.zip: Compressing text classification models. arXiv preprint arXiv:1612.03651.  
+Shima Khanehzar, Andrew Turpin, and Gosia Mikolajczak. 2019. Modeling political framing across policy issues and contexts. In Proceedings of the The 17th Annual Workshop of the Australasian Language Technology Association, ALTA 2019, Sydney, Australia, December 4-6, 2019, pages 61-66. Australasian Language Technology Association.  
+Klaus Krippendorff. 2011. Computing krippendorff's alpha-reliability.  
+Haewoon Kwak, Jisun An, and Yong-Yeol Ahn. 2020. A systematic media frame analysis of 1.5 million new york times articles from 2000 to 2017. In WebSci '20: 12th ACM Conference on Web Science, Southampton, UK, July 6-10, 2020, pages 305-314. ACM.  
+Vladimir I Levenshtein et al. 1966. Binary codes capable of correcting deletions, insertions, and reversals. In Soviet physics doklady, volume 10, pages 707-710. Soviet Union.  
+Siyi Liu, Lei Guo, Kate Mays, Margrit Betke, and Derry Tanti Wijaya. 2019. Detecting frames in news headlines and its application to analyzing news framing trends surrounding U.S. gun violence. In Proceedings of the 23rd Conference on Computational Natural Language Learning (CoNLL), pages 504-514, Hong Kong, China. Association for Computational Linguistics.  
+Julia Mendelsohn, Ceren Budak, and David Jurgens. 2021. Modeling framing in immigration discourse on social media. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2219-2263, Online. Association for Computational Linguistics.  
+Nona Naderi and Graeme Hirst. 2017. Classifying frames at the sentence level in news articles. In Proceedings of the International Conference Recent Advances in Natural Language Processing, RANLP 2017, pages 536-542, Varna, Bulgaria. INCOMA Ltd.  
+Viet-An Nguyen. 2015. Guided probabilistic topic models for agenda-setting and framing. Ph.D. thesis, University of Maryland, College Park.  
+OpenAI. 2023. Gpt-4 technical report. ArXiv, abs/2303.08774.  
+Guilherme Penedo, Quentin Malartic, Daniel Hesslow, Ruxandra Cojocaru, Alessandro Cappelli, Hamza Alobeidli, Baptiste Pannier, Ebtesam Almazrouei, and Julien Launay. 2023. The refinedweb dataset for falcon llm: Outperforming curated corpora with web data, and web data only.
+
+Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9.  
+Ricardo Rei, Nuno M Guerreiro, José Pombal, Daan van Stigt, Marcos Treviso, Luisa Coheur, José GC de Souza, and André FT Martins. 2023. Scaling up cometkiwi: Unbabel-ist 2023 submission for the quality estimation shared task. arXiv preprint arXiv:2309.11925.  
+Margaret E Roberts, Brandon M Stewart, Dustin Tingley, Christopher Lucas, Jetson Leder-Luis, Shana Kushner Gadarian, Bethany Albertson, and David G Rand. 2014. Structural topic models for open-ended survey responses. American journal of political science, 58(4):1064-1082.  
+Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. 2023. Llama 2: Open foundation and finetuned chat models.  
+Grigorios Tsoumakas and Ioannis Katakis. 2007. Multilabel classification: An overview. International Journal of Data Warehousing and Mining (IJDWM), 3(3):1-13.  
+Yuxiang Zhao and Qinghua Zhu. 2014. Evaluation on crowdsourcing research: Current status and future direction. Information systems frontiers, 16:417-434.  
+Liu Zhuang, Lin Wayne, Shi Ya, and Zhao Jun. 2021. A robustly optimized BERT pre-training approach with post-training. In Proceedings of the 20th Chinese National Conference on Computational Linguistics, pages 1218-1227, Huhhot, China. Chinese Information Processing Society of China.
+
+# A Annotation Schema
+
+We used a crowdsourcing approach with the help of non-expert annotators to create our training corpus, simplifying the process compared to the traditional method of hand-annotating by expert linguists and social science scholars, which is both expensive and inefficient. We collected data for the corpus in collaboration with graduate students whose task was to gather 150 sentences each, in various languages, from news articles related to the domains of immigration and same-sex marriage. These sentences were then annotated using the 15 framing dimensions established in the study (Boydstun et al., 2014), which are globally accepted, shown in Table 5.
+
+<table><tr><td>Frames</td><td>Definitions</td></tr><tr><td>Economic</td><td>The financial consequences and economic implications of the matter on various levels (person, family, community or broader economy).</td></tr><tr><td>Capacity and Resources</td><td>The presence or absence of various resources(physical, geographic, human, and financial) and the ability of existing systems.</td></tr><tr><td>Morality</td><td>Perspectives, policy objectives, or actions driven by religious principles, duties, ethics, or social responsibilities.</td></tr><tr><td>Fairness and Equality</td><td>The balance or distribution of laws, rights, and resources among individuals or groups.</td></tr><tr><td>Legality, Constitutionality, Jurisdiction</td><td>Discusses rights, freedoms and authority of individuals, corporations, and government.</td></tr><tr><td>Policy Prescription and Evaluation</td><td>Specific policies proposed to address identified issues and the assessment of policy effectiveness.</td></tr><tr><td>Crime and Punishment</td><td>Effectiveness and implications of laws and their enforcement.</td></tr><tr><td>Security and Defense</td><td>Actions or calls to action aimed at protecting individuals, groups, or nations from potential threats to their well-being.</td></tr><tr><td>Health and Safety</td><td>Access to healthcare, health outcomes, disease, sanitation, mental health, violence prevention, infrastructure safety, and public health.</td></tr><tr><td>Quality of life</td><td>Threats and opportunities for the individual&#x27;s wealth, happiness and well being.</td></tr><tr><td>Cultural Identity</td><td>Traditions, customs or values of a social group in relation to a policy issue.</td></tr><tr><td>Public Sentiment</td><td>References of attitudes and opinions of the general public, including polling and demographics.</td></tr><tr><td>Political</td><td>Political considerations, actions, efforts, stances, and partisan, bipartisan, or lobbying activities related to an issue.</td></tr><tr><td>External Regulation and Rep- ution</td><td>The external relations of nations or groups, trade agreements, policy outcomes, and external perceptions or consequences.</td></tr><tr><td>Other</td><td>Frames that don&#x27;t fit into the categories above.</td></tr></table>
+
+Table 5: Frames and their definitions as outlined by Policy Frames Codebook (PFC, Boydstun et al. (2014)). This codebook was given to the students as annotation schema.
+
+B Novel Bengali and Portuguese Test Set Statistic  
+
+<table><tr><td>Number of sentences</td><td>Bengali</td><td>Portuguese</td></tr><tr><td>Economic</td><td>36</td><td>20</td></tr><tr><td>Capacity and Resources</td><td>3</td><td>19</td></tr><tr><td>Morality</td><td>4</td><td>13</td></tr><tr><td>Fairness and Equality</td><td>13</td><td>23</td></tr><tr><td>Legality Constitutional-ity Jurisdiction</td><td>12</td><td>25</td></tr><tr><td>Policy Prescription and Evaluation</td><td>13</td><td>24</td></tr><tr><td>Crime and Punishment</td><td>11</td><td>3</td></tr><tr><td>Security and Defence</td><td>5</td><td>23</td></tr><tr><td>Health and Safety</td><td>14</td><td>9</td></tr><tr><td>Quality of Life</td><td>33</td><td>15</td></tr><tr><td>Cultural Identity</td><td>1</td><td>32</td></tr><tr><td>Public Sentiment</td><td>5</td><td>24</td></tr><tr><td>Political</td><td>3</td><td>10</td></tr><tr><td>External Regulation and Reputation</td><td>41</td><td>1</td></tr><tr><td>Other</td><td>34</td><td>3</td></tr><tr><td>Total</td><td>228</td><td>244</td></tr></table>
+
+Table 6: Number of texts per frame per language
+
+The distribution of labels in the Bengali and Portuguese test sets (see Table 6) reveals intriguing domain affinity. In the case of Bengali, the news articles predominantly focus on the immigration domain, reflecting the cultural disparities between Brazil and Bangladesh. Specifically, the test set emphasizes the economic and lifestyle aspects of immigration (Bengali), while also delving into the legal and policy-making dimensions of the domain (Portuguese).
+
+# C Assessing Translation Quality
+
+Table 7 shows the breakdown of the comet score per language.  
+
+<table><tr><td>Language Pair</td><td>Comet Score (%)</td></tr><tr><td>English-Bengali</td><td>74.39</td></tr><tr><td>English-German</td><td>76.93</td></tr><tr><td>English-Greek</td><td>76.64</td></tr><tr><td>English-Hindi</td><td>67.87</td></tr><tr><td>English-Italian</td><td>79.04</td></tr><tr><td>English-Nepali</td><td>86.84</td></tr><tr><td>English-Russian</td><td>79.87</td></tr><tr><td>English-Swahili</td><td>73.71</td></tr><tr><td>English-Telugu</td><td>69.02</td></tr><tr><td>English-Bengali</td><td>78.79</td></tr><tr><td>English-Turkish</td><td>74.63</td></tr><tr><td>English-Chinese</td><td>74.63</td></tr><tr><td>English-Portuguese</td><td>74.89</td></tr><tr><td>System Score</td><td>76.05</td></tr></table>
+
+Table 7: Average score from CometWiki of the Automatic Translation Quality without reference. The high resource languages (i.e., Italian, Greek etc) have higher scores than lower resource languages (i.e., Telugu)
+
+# D Complete Results for English and Multilingual Experiments
+
+We observed the mean accuracy of the MFC evaluation set for models trained on English and Multilingual datasets. The key findings are summarized below:
+
+1. The MFC alone achieved higher accuracy compared to other systems, with scores of  $61.93\%$  and  $69.52\%$  for BERT and RoBERTa-based models, respectively. However, when using the MFC10 dataset with limited high-quality data, the accuracy dropped significantly to  $53.02\%$  and  $57.45\%$  for BERT and RoBERTa models, respectively.  
+2. The SNFC and MaSNFC datasets exhibited lower accuracy when evaluated individually, compared to the MFC. However, the SNFC outperformed MFC10 in terms of accuracy for the BERT model. The SNFC has an accuracy of  $60.57\%$  while the MFC10 has gotten  $53.02\%$ . It is worth noting that the larger size of the SNFC contributed to its higher accuracy compared to MaSNFC, which is almost three times smaller.  
+3. Combining the MFC with our datasets led to substantial accuracy improvements. The models trained on MFC+SNFC (72.57%, 72.07%) and MFC+MaSNFC (72.85%, 73.22%) achieved higher accuracy than the MFC alone (61.93%, 69.52%), for both BERT and RoBERTa models.  
+4. Combining MFC10 with our datasets, we observed improved accuracy as well. The MFC10+SNFC combination yielded an accuracy improvement of 6.1 and 4.77 percentage points for BERT and RoBERTa models, respectively, compared to MFC10. Similarly, MFC10+MaSNFC demonstrated a similar improvement of 7.1 and 3.49 percentage points, respectively.  
+5. The overall accuracies of the MFC evaluation set for multilingual data (Table 3) are lower compared to the accuracies for English training (Table 2). This can be attributed to the fact that the training data in other languages were obtained through automatic translation, which may not be of the same quality as human translations or original news articles in those languages.
+
+<table><tr><td>System Name</td><td>Number of Sentences</td><td>BERT</td><td>RoBERTa</td></tr><tr><td>MFC</td><td>9740</td><td>61.93</td><td>69.52</td></tr><tr><td>MFC10</td><td>1125</td><td>53.02</td><td>57.45</td></tr><tr><td>SNFC</td><td>17520</td><td>60.57</td><td>54.37</td></tr><tr><td>MaSNFC</td><td>5182</td><td>52.05</td><td>48.77</td></tr><tr><td>MFC+</td><td>27260</td><td>72.57</td><td>72.07</td></tr><tr><td>SNFC</td><td></td><td></td><td></td></tr><tr><td>MFC+</td><td>14922</td><td>72.85</td><td>73.22</td></tr><tr><td>MaSNFC</td><td></td><td></td><td></td></tr><tr><td>MFC10+</td><td>18645</td><td>68.03</td><td>64.75</td></tr><tr><td>SNFC</td><td></td><td></td><td></td></tr><tr><td>MFC10+</td><td>6307</td><td>60.12</td><td>60.94</td></tr><tr><td>MaSNFC</td><td></td><td></td><td></td></tr></table>
+
+Table 8: Mean Accuracy Scores on the MFC evaluation set for models trained on English Datasets. The best scores have been highlighted.
+
+6. Among the datasets, MFC+MaSNFC achieved the highest accuracy of 45.73 on the multilingual test set, outperforming both MFC and MFC10 datasets.  
+7. For the Bengali test set, the highest accuracy (32.02) was achieved by the MFC10+SNFC training dataset. As for the Portuguese test set, the highest accuracy of 33.61 was obtained by two systems: MFC and MFC+MaSNFC.  
+8. Overall, the accuracies for the Bengali and Portuguese test sets were lower than those for the MFC evaluation set. This can be attributed to two factors. First, the training data, being translations, may not capture the nuances of the original news articles. Second, the training data mainly consists of MFC, which is collected from US-based news media sources. The test sets, on the other hand, were collected from Brazil and Bangladesh, which have different cultural contexts in their news articles that cannot be fully replicated through translation. To improve the scores further, it would be necessary to obtain original news articles from diverse culturally distinct sources in different languages.
+
+The study highlights challenges in multilingual framing analysis, with lower accuracies compared to English training. It emphasizes the need for high-quality translations and original news articles. Combining datasets like MFC+MaSNFC can enhance accuracy. Considering cultural and linguistic con
+
+<table><tr><td rowspan="2">System Name</td><td colspan="2">MFC Evaluation Set</td><td colspan="2">Bengali Test Set</td><td colspan="2">Portuguese Test Set</td></tr><tr><td>mBERT</td><td>XLM-R</td><td>mBERT</td><td>XLM-R</td><td>mBERT</td><td>XLM-R</td></tr><tr><td>MFC (English)</td><td>27.70</td><td>28.13</td><td>16.67</td><td>25.44</td><td>26.23</td><td>28.28</td></tr><tr><td>MFC</td><td>44.87</td><td>44.99</td><td>21.93</td><td>25.88</td><td>30.33</td><td>33.61</td></tr><tr><td>MFC10</td><td>27.7</td><td>28.64</td><td>20.61</td><td>23.68</td><td>30.33</td><td>27.87</td></tr><tr><td>SNFC</td><td>28.05</td><td>28.04</td><td>22.37</td><td>25.44</td><td>27.05</td><td>23.77</td></tr><tr><td>MaSNFC</td><td>28.86</td><td>27.55</td><td>11.84</td><td>16.67</td><td>20.49</td><td>15.98</td></tr><tr><td>MFC+SNFC</td><td>45.09</td><td>44.07</td><td>23.25</td><td>26.31</td><td>29.92</td><td>31.56</td></tr><tr><td>MFC+MaSNFC</td><td>44.42</td><td>45.73</td><td>22.37</td><td>28.07</td><td>31.97</td><td>33.61</td></tr><tr><td>MFC10 + SNFC</td><td>30.01</td><td>33.11</td><td>25</td><td>32.02</td><td>29.51</td><td>26.62</td></tr><tr><td>MFC10+MaSNFC</td><td>33.33</td><td>32.56</td><td>22.81</td><td>24.56</td><td>22.13</td><td>26.64</td></tr></table>
+
+Table 9: Mean Accuracy Scores on the MFC evaluation set and Novel Multilingual Test Set for models trained on Multilingual Datasets. The best scores have been highlighted.
+
+texts and diverse training data is crucial for better understanding framing across languages and cultures.
+
+# E Instruction for the Generative AI Models
+
+This was the instruction that was given to the models discussed in Section 5.
+
+"In this task, you will be provided with a list of frames and a sentence. Your goal is to select the single most suitable frame from the given list for the provided sentence. Frames are cognitive structures that help humans interpret information by providing a mental framework for understanding. Each frame represents a specific perspective, context, or interpretation. Frame Selection Format: In your response, do not write anything other than the name of the frame. Frames List and Definitions: 'Economic': 'The financial consequences and economic implications of the matter on various levels (person, family, community or broader economy)'. 'External Regulation and Reputation': 'The external relations of nations or groups, trade agreements, policy outcomes, and external perceptions or consequences.' 'Political': 'Political considerations, actions, efforts, stances, and partisan, bipartisan, or lobbying activities related to an issue.' 'Public Sentiment': 'References of attitudes and opinions of the general public, including polling and demographics.' 'Cultural Identity': 'Traditions, customs, or values of a social group in relation to a policy issue.' 'Quality of Life': 'Threats and opportunities for the individual's wealth, happiness, and well-being.' 'Health and Safety': 'Access to healthcare, health outcomes, disease, sanitation, mental health, violence prevention, infrastructure safety, and public health.' 'Security and Defense': 'Actions or calls to action aimed at protecting individuals, groups, or nations from potential threats to their well-being.' 'Crime and Punishment': 'Effectiveness and implications of laws and their enforcement.' 'Policy Prescription and Evaluation': 'Specific policies proposed to address identified issues and the assessment of policy effectiveness.' 'Legality, Constitutionality, Jurisdiction': 'Discusses rights, freedoms, and authority of individuals, corporations, and government.' 'Fairness and Equality': 'The balance or distribution of laws, rights, and resources among individuals or groups.' 'Morality': 'Perspectives, policy objectives, or actions driven by religious principles, duties, ethics, or social responsibilities.' 'Capacity and Resources': 'The presence or
+
+absence of various resources (physical, geographic, human, and financial) and the ability of existing systems.,'Other': 'Frames that don't fit into the categories above.' Please select the most appropriate frame for the given sentence, and specify the chosen frame without additional commentary. You are allowed to choose only one frame from the provided list that best aligns with the meaning and context of the given sentence. Consider the nuances of the sentence and the various frames in the list before making your selection. Your choice should reflect the frame that you believe most accurately captures the intended interpretation of the sentence. Remember that frames can significantly impact how information is perceived and understood. Choose the frame that enhances the sentence's meaning and effectively shapes the interpretation. Your ability to accurately match sentences with appropriate frames will contribute to successful communication and understanding. Do not write anything like, 'The selected frame is.' Input: "
+
+# F Erroneous Predictions from Generative LLMs
+
+<table><tr><td>Text</td><td>Gold Label Output</td><td>Models</td><td>Model Prediction</td></tr><tr><td>DETENTION OF REFUGEEES IN TEXAS OK&#x27;D</td><td>Crime and Punishment</td><td>Llama 70B</td><td>The External Regulations And reputation was picked since it refers nation borders control migrants entering its territory as refugee status;</td></tr><tr><td>But last spring, the attorney for the regents, Burns Newsome, advised them to end the waivers for illegal immigrants, saying the practice is at odds with SB 529. Now, the regents are having a series of public hearings to get input on the issue.</td><td>Legality, Constitu-tionality, Jurisdiction</td><td>Llama 70B</td><td>External Regulation and Renown</td></tr><tr><td>3 Prison Guards Guilty of Abuse Of Immigrants</td><td>Crime and Punishment</td><td>Falcon 40B</td><td>Prison guards guilty.</td></tr><tr><td>DHS budget brinkmanship could leave local governments in the lurch</td><td>Economic</td><td>Falcon 40B</td><td>Economical</td></tr><tr><td>Disorder On The Border</td><td>Security and Defense</td><td>Mistral 7B</td><td>The most suitable frame for the sentence ‘Disorder on the Border’ is ‘Security and Defense’</td></tr><tr><td>Police say Orleans-Lindsay shot Toatley, 37, to death Oct. 30 during an undercover drug purchase. Orleans-Lindsay, of Silver Spring, allegedly accepted $3,500 in cash for cocaine, pretended to get the drugs and returned to the trooper&#x27;s sport-utility vehicle a few minutes later, firing a single shot. Toatley, a 16-year trooper and father of three, died 2 1/2 hours later.</td><td>Crime and Punishment</td><td>Mistral 7B</td><td>‘Law Enforcement’, ‘Political’, ‘Fair-ness and Equality’, ‘Crime and Punishment’ and ‘Legality, Constitutionality, Jurisdiction’</td></tr></table>
+
+Table 10: LLMs predictions compared to gold label outputs. Errors include spelling errors to not following instruction.

2024/A Study on Scaling Up Multilingual News Framing Analysis/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c9a5b5f48b9ce03796259acb806d5f7153d8958d43acdc080cc7141151e2a57d
+size 927258

2024/A Systematic Analysis of Subwords and Cross-Lingual Transfer in Multilingual Translation/d12825ff-73dc-4d7f-b285-a37a03d24322_content_list.json

@@ -0,0 +1,1034 @@
+[
+    {
+        "type": "text",
+        "text": "A Systematic Analysis of Subwords and Cross-Linguual Transfer in Multilingual Translation",
+        "text_level": 1,
+        "bbox": [
+            231,
+            90,
+            764,
+            130
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Francois Meyer and Jan Buys",
+        "bbox": [
+            364,
+            152,
+            636,
+            168
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Department of Computer Science",
+        "bbox": [
+            364,
+            168,
+            636,
+            185
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "University of Cape Town",
+        "bbox": [
+            398,
+            186,
+            603,
+            202
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "francois.meyer@uct.ac.za, jbuys@cs.uct.ac.za",
+        "bbox": [
+            278,
+            203,
+            722,
+            218
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            260,
+            339,
+            275
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Multilingual modelling can improve machine translation for low-resource languages, partly through shared subword representations. This paper studies the role of subword segmentation in cross-lingual transfer. We systematically compare the efficacy of several subword methods in promoting synergy and preventing interference across different linguistic typologies. Our findings show that subword regularisation boosts synergy in multilingual modelling, whereas BPE more effectively facilitates transfer during cross-lingual fine-tuning. Notably, our results suggest that differences in orthographic word boundary conventions (the morphological granularity of written words) may impede cross-lingual transfer more significantly than linguistic unrelatedness. Our study confirms that decisions around subword modelling can be key to optimising the benefits of multilingual modelling.",
+        "bbox": [
+            144,
+            288,
+            460,
+            575
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            588,
+            258,
+            602
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Machine translation (MT) models have become increasingly multilingual (Dabre et al., 2020). This greatly benefits low-resource languages through positive transfer from high-resource languages (Ha et al., 2016; Aharoni et al., 2019). However, increasing multilinguality in a limited shared parameter space can lead to suboptimal performance for high-resource languages (Firat et al., 2016; Team et al., 2022). There is a tradeoff between maximising positive cross-lingual transfer (also known as synergy) while minimising negative cross-lingual interaction (also known as interference).",
+        "bbox": [
+            112,
+            614,
+            489,
+            806
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Several modelling decisions affect synergy and interference in multilingual MT. Shaham et al. (2023) experimentally analysed the influence of factors like model size and language data proportions. One aspect their study failed to consider is subword segmentation. The shared subword vocabulary of multilingual models presents a similar",
+        "bbox": [
+            112,
+            808,
+            489,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/c3659bb69add3615263b79c8fbafb932be5dff55dabf9eaa7f9442b17ae84bea.jpg",
+        "image_caption": [
+            "Figure 1: Performance increase for English  $\\rightarrow$  Siswati through multilingual modelling varies greatly across subword methods and linguistic contexts."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            510,
+            260,
+            873,
+            425
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "trade-off as the shared parameter space - overlapping subword representations induce synergy, but having to represent multiple languages in a limited vocabulary can harm cross-lingual transfer (Chung et al., 2020; Rust et al., 2021; Patil et al., 2022).",
+        "bbox": [
+            507,
+            512,
+            884,
+            592
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In this paper we experimentally analyse the role of subword segmentation in multilingual and cross-lingual MT. Our goal is to compare different classes of subword methods with regards to their ability to induce synergy, reduce interference, and transfer knowledge during cross-lingual finetuning. We also investigate how cross-lingual transfer is influenced by the linguistic similarities of interacting languages, with particular focus on factors related to subword structure like morphological typology and orthographic word boundary conventions (the degree to which morphemes are concatenated or written as separate orthographic words).",
+        "bbox": [
+            507,
+            596,
+            884,
+            804
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "We run experiments on translation from English to four linguistically diverse South African languages (see Table 1). This selection covers different levels of language relatedness, morphological complexity, and orthographic word granularity, allowing us to analyse how these factors interact with different subword methods to influence cross-",
+        "bbox": [
+            507,
+            808,
+            885,
+            920
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "2194",
+        "bbox": [
+            478,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: NAACL 2024, pages 2194-2200",
+        "bbox": [
+            218,
+            945,
+            778,
+            957
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "June 16-21, 2024 ©2024 Association for Computational Linguistics",
+        "bbox": [
+            292,
+            957,
+            704,
+            971
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "table",
+        "img_path": "images/f6effd8a215cbc6ef939ab8f383e6c337604c09f4447be5a9370a082b14c8ec4.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Language</td><td>Family</td><td>Morphology</td><td>Orthography</td><td>What is your name?</td><td>Thank you!</td></tr><tr><td>Siswati (ss)</td><td>NC/Bantu/Nguni</td><td>agglutinative</td><td>conjunctive</td><td>Ngubani ligama lakho?</td><td>Ngiyabonga!</td></tr><tr><td>isiXhosa (xh)</td><td>NC/Bantu/Nguni</td><td>agglutinative</td><td>conjunctive</td><td>Ungubani igama lakho?</td><td>Enkosi!</td></tr><tr><td>Setswana (ts)</td><td>NC/Bantu/Sotho-Tswana</td><td>agglutinative</td><td>disjunctive</td><td>Leina la gago ke mang?</td><td>Ke a leboga!</td></tr><tr><td>Afrikaans (af)</td><td>Indo-European/Germanic</td><td>analytic</td><td>disjunctive</td><td>Wat is jou naam?</td><td>Dankie!</td></tr></table>",
+        "bbox": [
+            115,
+            80,
+            882,
+            166
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Table 1: We vary the language modelled alongside Siswati to control relatedness, morphology, and orthography.",
+        "bbox": [
+            119,
+            174,
+            875,
+            190
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "lingual transfer. Low-resource languages stand to benefit most from multilingual modelling. In all our experiments we focus on cross-lingual transfer to Siswati, which is by far the least resourced among the languages included. It presents exactly the type of real world low-resource translation scenario we are interested in studying.",
+        "bbox": [
+            112,
+            215,
+            490,
+            326
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We conduct two sets of experiments - multilingual MT and cross-lingual finetuning. Our multilingual experiments follow Shaham et al. (2023) in training several trilingual MT models and comparing synergy/interference (see Figure 1). In the cross-lingual finetuning experiments we finetune pretrained bilingual MT models on new languages. Our results demonstrate that decisions around subword segmentation significantly affect MT performance. ULM (Kudo, 2018) improves synergy in multilingual modelling, while BPE (Sennrich et al., 2016) enhances cross-lingual transfer during finetuning. Going beyond linguistic relatedness, we find that the much less studied influence of orthographic word boundary conventions can drastically affect the cross-lingual transfer achieved between interacting languages.",
+        "bbox": [
+            115,
+            329,
+            489,
+            602
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            112,
+            618,
+            270,
+            633
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Synergy and interference are well-established phenomena (Firat et al., 2016; Aharoni et al., 2019; Team et al., 2022), but not well understood. Shatham et al. (2023) address this by systematically analysing the role of several factors in synergy and interference: (1) model size, (2) data size, (3) language proportions, (4) number of languages, and (5) language relatedness. Their results show that scaling model size and tuning the data sampling temperature greatly alleviates interference. They do not vary subword segmentation in their experiments, using the same sentencepiece (Kudo and Richardson, 2018) vocabulary across all models.",
+        "bbox": [
+            112,
+            646,
+            489,
+            854
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "However, multilingual vocabularies are known to affect cross-lingual transfer through factors such as cross-lingual subword overlap (Pires et al., 2019; Wu and Dredze, 2019; Patil et al., 2022) and under",
+        "bbox": [
+            112,
+            857,
+            489,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "represented low-resource languages (Wang et al., 2021; Acs, 2019). These issues have mainly been studied for multilingual language modelling (Rust et al., 2021; Maronikolakis et al., 2021; Chung et al., 2020), but the same concerns hold for MT (Wang et al., 2020a). We are unaware of existing work comparing the multitude of proposed subword methods in the context of multilingual MT.",
+        "bbox": [
+            507,
+            215,
+            884,
+            344
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3 Methodology",
+        "text_level": 1,
+        "bbox": [
+            507,
+            359,
+            658,
+            376
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "This study involves two sets of MT experiments - (1) multilingual (trilingual) experiments to investigate synergy/interference, and (2) finetuning experiments to analyse cross-lingual transfer. Our goal is to determine which subwords benefit low-resource languages and how cross-lingual transfer depends on linguistic typology. The linguistic diversity of South Africa is an ideal testing ground for our purposes. Siswati is a low-resource agglutinative language, so effective subword modelling is critical for dealing with the inevitably high proportion of out-of-vocabulary words in held-out datasets. We use Siswati as the low-resource target language in our experiments and alternate the higher resourced language modelled alongside Siswati between isi-Xhosa, Setsswana, and Afrikaans.",
+        "bbox": [
+            507,
+            387,
+            884,
+            643
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Table 1 shows how these language present varying linguistic relationships to Siswati. IsiXhosa is closely related. Setswana is somewhat related and also agglutinative, but diverges in its orthography - its writing system is disjunctive (Pretorius et al., 2009). This refers to how linguistic words (e.g. nouns, verbs) are represented as orthographic words (space-separated tokens). Disjunctive orthographies write a single linguistic word as multiple orthographic words (e.g. in Setswana prefixal morphemes are space-separated from verbal roots).",
+        "bbox": [
+            507,
+            646,
+            885,
+            822
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "While linguistic relatedness and morphological complexity are obvious features to consider in any analysis of cross-lingual interactions, we are unaware of work considering the impact of orthographic word boundary conventions. We include it as a factor in our study because of its potential",
+        "bbox": [
+            507,
+            824,
+            884,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "2195",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "relevance to subword segmentation. Orthographic word boundaries determine the pre-tokenisation of text before subword segmenters are applied, so it could well affect aspects like segmentation granularity and overlap between the subword vocabularies of different languages. Afrikaans is linguistically unrelated to Siswati and also disjunctive, but because of its analytic morphology (lower morpheme-to-word ratio) its written words are sometimes more aligned to those of Siswati (e.g. see phrases in Table 1).",
+        "bbox": [
+            110,
+            84,
+            489,
+            261
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "This selection of languages allows us to isolate the cross-lingual effects of linguistic relatedness, morphological typology, and orthographic word boundary conventions. In the case of Setswana-Siswati, we can study whether the potentially positive cross-lingual effect of their linguistic relatedness is nullified by the fact that the two languages have very different conventions for orthographic word boundaries.",
+        "bbox": [
+            110,
+            263,
+            489,
+            409
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Multilingual Modelling",
+        "text_level": 1,
+        "bbox": [
+            112,
+            430,
+            344,
+            448
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We train two bilingual models and one trilingual model per language pair (see Table 2).",
+        "bbox": [
+            112,
+            458,
+            487,
+            491
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "table",
+        "img_path": "images/faf463cfbaf5992c3cad09c31352e08c9bd0cd5ab6d3ad66552131c9d17f0c3d.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Languages</td><td>Examples</td><td>Subwords</td></tr><tr><td>en→ss</td><td>166k</td><td>BPE/ULM/SSMT</td></tr><tr><td>en→xh/ts/af</td><td>1.6m</td><td>BPE/ULM/SSMT</td></tr><tr><td>en→ss+ xh/ts/af</td><td>1.6m+166k</td><td>BPE/ULM/SSMT/OBPE</td></tr></table>",
+        "bbox": [
+            117,
+            507,
+            485,
+            579
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Multilingual experimental setup: bilingual and trilingual models (bilingual OBPE is equivalent to BPE).",
+        "bbox": [
+            112,
+            589,
+            485,
+            619
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "This setup allows us to compare how MT performance changes for en  $\\rightarrow$  ss and en  $\\rightarrow$  xh/ts/af going from bilingual models to multilingual models. Following Shaham et al. (2023), we measure synergy/interference for a translation direction  $s\\to t$  by the relative difference in performance between a bilingual model trained to translate only from  $s$  to  $t$  and a multilingual model trained to translate to an additional language.",
+        "bbox": [
+            110,
+            644,
+            489,
+            788
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Shaham et al. (2023) use test set cross-entropy loss to measure MT performance, but this cannot be reliably used to compare across different subword segmentations. Instead, we use test set  $\\mathrm{chrF}++$  (Popovic, 2017) to measure performance. It is a reference-based metric that combines word and character information, so it is well suited for evaluating subword-level performance. Our modified",
+        "bbox": [
+            110,
+            791,
+            489,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "formula for measuring synergy/interference is",
+        "bbox": [
+            507,
+            84,
+            853,
+            101
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {I} _ {s \\rightarrow t} = \\frac {\\mathrm {C H R F} + + (M _ {s \\rightarrow t} ^ {\\mathrm {m u l t i}}) - \\mathrm {C H R F} + + (M _ {s \\rightarrow t} ^ {\\mathrm {b i}})}{\\mathrm {C H R F} + + (M _ {s \\rightarrow t} ^ {\\mathrm {b i}})},\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            519,
+            111,
+            870,
+            149
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $M$  are trained multilingual/bilingual models evaluated on  $s\\rightarrow t$  translation. Negative values of  $I_{s\\rightarrow t}$  indicate worse performance for  $s\\rightarrow t$  in the multilingual model (interference) and positive values indicate improved performance (synergy).",
+        "bbox": [
+            507,
+            158,
+            882,
+            239
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2 Cross-Lingual Finetuning",
+        "text_level": 1,
+        "bbox": [
+            507,
+            250,
+            759,
+            266
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We train a bilingual subword segmenter and MT model for en→xh/ts/af, and then finetune and evaluate the model in the other translation directions (e.g. pretrain en→xh and finetune on en→ss, en→ts, and en→af). Varying the subword method reveals how different subwords facilitate crosslingual transfer during finetuning from higher resourced languages (isiXhosa/Setswana/Afrikaans) to lower resourced Siswati.",
+        "bbox": [
+            507,
+            271,
+            884,
+            414
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4 Experimental Setup",
+        "text_level": 1,
+        "bbox": [
+            507,
+            426,
+            717,
+            445
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We compare five subword segmenters (four per experiment). We chose methods that represent the main paradigms of subword segmentation - deterministic segmentation, subword regularisation, learning subwords during training, and subword techniques for enhancing cross-lingual transfer.",
+        "bbox": [
+            507,
+            453,
+            882,
+            549
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. BPE (Sennrich et al., 2016) iteratively adds frequently co-occurring subwords to its vocabulary. We use it as a deterministic segmenter.",
+            "2. ULM (Kudo, 2018) learns segmentation to optimise a unigram language model and can be used as a probabilistic segmenter, exposing models to multiple subword segmentations for regularisation. We set the sampling parameter  $\\alpha$  to 0.5.",
+            "3. SSMT (Meyer and Buys, 2023) is a subword segmental MT model which learns subword segmentation jointly during MT training, with the goal of learning subwords that optimise MT performance.",
+            "4. OBPE (Patil et al., 2022) modifies BPE to boost subword overlap among languages in multilingual vocabularies. We use it in our multilingual experiments to see if increased shared subword representations promote synergy.",
+            "5. XBPE (Wang et al., 2020b) extends the BPE vocabulary of a pretrained model to include BPE subwords of a new translation direction. New subword embeddings are randomly initialised. We use it in our finetuning experiments to see if the vocabulary extension enhances cross-lingual transfer."
+        ],
+        "bbox": [
+            507,
+            550,
+            885,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "2196",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/fe341bc27bd12b8ea9a8921aaf6c92d34d1f5543463deb49e678a0af795d87b7.jpg",
+        "image_caption": [
+            "Figure 2: Performance change for en→xh/af/ts through multilingual modelling alongside en→ss."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            115,
+            82,
+            480,
+            211
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/4d68c985bcc03a71dd3f0041e0375ea5d8709b679042128aca2c651bd9a2a5a1.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>tgt</td><td>BPE</td><td>ULM</td><td>SSMT</td><td>OBPE</td></tr><tr><td rowspan=\"2\">en→ss/xh</td><td>ss</td><td>33.4</td><td>35.1</td><td>33.6</td><td>31.1</td></tr><tr><td>xh</td><td>46.8</td><td>47.2</td><td>46.1</td><td>44.6</td></tr><tr><td rowspan=\"2\">en→ss/af</td><td>ss</td><td>28.2</td><td>30.6</td><td>29.7</td><td>27.4</td></tr><tr><td>af</td><td>60.9</td><td>61.1</td><td>60.2</td><td>60.2</td></tr><tr><td rowspan=\"2\">en→ss/ts</td><td>ss</td><td>22.5</td><td>27.5</td><td>17.8</td><td>18.8</td></tr><tr><td>ts</td><td>31.3</td><td>34.7</td><td>23.6</td><td>26.4</td></tr></table>",
+        "bbox": [
+            122,
+            263,
+            480,
+            379
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Test set  $\\mathrm{{chrF}} +  +$  of trilingual models.",
+        "bbox": [
+            144,
+            388,
+            453,
+            403
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Training We train models on WMT22 data (Adelani et al., 2022) and validate and test on FLORES (Goyal et al., 2021; Team et al., 2022). The number of training sentences are shown in Table 2. For en  $\\rightarrow$  ss this is the full WMT22 dataset, but for en  $\\rightarrow$  xh/ts/af we sampled sentences from en  $\\rightarrow$  xh and en  $\\rightarrow$  ts to match the size of en  $\\rightarrow$  af, removing data size as an influence. We also removed examples from en  $\\rightarrow$  xh/ts/af where English source sentences were found in en  $\\rightarrow$  ss to neutralise the positive transfer effect of multi-parallel overlap (Stap et al., 2023). The hyperparameters of our models and subword methods are included in Appendix A.",
+        "bbox": [
+            112,
+            429,
+            489,
+            639
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "5 Results & Discussion",
+        "text_level": 1,
+        "bbox": [
+            112,
+            651,
+            329,
+            668
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We plot the synergy/interference analysis of our multilingual experiments in Figures 1 & 2, while the absolute performance of the models are provided in Table 3. The results from our cross-lingual finetuning experiments are presented in Figure 3.",
+        "bbox": [
+            112,
+            677,
+            487,
+            759
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "5.1 Which subwords promote synergy and minimise interference?",
+        "text_level": 1,
+        "bbox": [
+            112,
+            771,
+            465,
+            802
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "ULM consistently achieves greater synergy than other subword methods. This holds across all linguistic contexts (Fig. 1) and results in better absolute performance in all translation directions (Table 3). It comes at the cost of minimal interference for the higher resourced languages, and even some synergy for en→ts (Fig. 2). The subword regulari",
+        "bbox": [
+            112,
+            808,
+            489,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/896c6c07ad2f3f3d7b46abd1f15f9fc6aa762874dbf3403014fd0282d0267c29.jpg",
+        "image_caption": [
+            "Figure 3: Test set  $\\mathrm{chrF}++$  of pretraining for en  $\\rightarrow$  xh/af/ts (rows) and finetuning on en  $\\rightarrow$  xh/af/ts/ss (columns). Diagonal entries are bilingual models with no finetuning."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            509,
+            80,
+            873,
+            292
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "sation of ULM ensures that models are more robust to the varied subwords of multilingual modelling.",
+        "bbox": [
+            507,
+            370,
+            880,
+            403
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "5.2 Which subwords transfer cross-lingually?",
+        "text_level": 1,
+        "bbox": [
+            507,
+            414,
+            880,
+            430
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "BPE subwords exhibit the greatest cross-lingual transferability. In contrast to our multilingual findings, the subword regularisation of ULM proves a barrier to cross-lingual finetuning. ULM is a probabilistic segmenter that is sampled during training, but when the probabilistic model is based on one language and applied to another, its samples might yield highly inadequate subword units. The consistent deterministic segmentation of BPE allows the finetuned model to adapt to a new translation direction effectively.",
+        "bbox": [
+            505,
+            436,
+            882,
+            613
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "5.3 What is the role of linguistic typology?",
+        "text_level": 1,
+        "bbox": [
+            507,
+            625,
+            857,
+            642
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "A consistent pattern emerges in the cross-lingual dynamics between Siswati and other languages. IsiXhosa modelling proves to be most beneficial for Siswati performance. Afrikaans achieves less transfer, presumably because it is not related. Somewhat surprisingly, the weakest synergy is between Siswati and Setswana, even though both are agglutinative Bantu languages. This highlights the impact of orthographic systems on cross-lingual transfer: diverging word boundary conventions can impede cross-lingual transfer more than linguistic unrelatedness. Data-driven multilingual models that learn from text might miss underlying similarities between languages that are obscured by superficial differences in their surface realisations.",
+        "bbox": [
+            505,
+            646,
+            882,
+            887
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Our results highlight two interacting effects. Firstly, linguistic relatedness does play a role -",
+        "bbox": [
+            507,
+            889,
+            884,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "2197",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "isiXhosa consistently improves Siswati more than Setsswana and Afrikaans. Secondly, in the specific case of Setsswana-Siswati, their relatedness is rendered all but irrelevant by the fact that the languages have diverging orthographies. Afrikaans does not have the extremely disjunctive orthography of Setsswana so even though it is less related to Siswati than Setsswana, the orthography of Setsswana prevents transfer to Siswati. Linguistic distance plays a role in both cases (Afrikaans-Siswati and Setsswana-Siswati) but for Setsswana-Siswati it is a less important factor than orthography.",
+        "bbox": [
+            112,
+            84,
+            492,
+            294
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Orthography is a notable difference between Nguni languages like Siswati and Sotho-Tswana languages like Setsswana (Pretorius et al., 2009). Taljard and Bosch (2006) showed that the diverging orthographies of these two language groups necessitate different approaches to more traditional NLP tasks, even though the languages are linguistically and morphologically related. Our results suggest a similar situation for cross-lingual transfer in multilingual modelling: Differences in orthographic word boundary conventions harms synergy between otherwise related languages.",
+        "bbox": [
+            115,
+            294,
+            490,
+            488
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "6 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            112,
+            500,
+            247,
+            514
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We presented an in-depth study on the role of subwords in multilingual and cross-lingual MT. Our results demonstrate that subword segmentation significantly influences cross-lingual interactions. ULM proves optimal for transfer to low-resource languages in multilingual modelling, while BPE enables greater cross-lingual transfer during finetuning. Besides language relatedness, we show that similarities/differences in orthographic word granularity can greatly affect multilingual performance. There is more work to be done on the role of orthographic word boundary conventions in neural MT. Future work could aim to design multilingual techniques that see past orthographic differences in order to leverage more fundamental similarities between languages.",
+        "bbox": [
+            112,
+            527,
+            489,
+            785
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            112,
+            797,
+            220,
+            813
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Our study is limited to translation from English to four South African languages. While the chosen languages are typologically diverse, our conclusions might not necessarily hold for languages from different language families and with distinct orthographies. We did not consider languages that",
+        "bbox": [
+            112,
+            824,
+            489,
+            921
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "have multiple orthographies, which might be another approach to study the effects of orthography. The performance differences between different subword segmentation methods across languages in our results are relatively consistent, but a more detailed analysis on the interaction between choice of subword segmentation method and language could yield additional explanations of the results.",
+        "bbox": [
+            507,
+            84,
+            885,
+            212
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            509,
+            225,
+            682,
+            240
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "This work is based on research supported in part by the National Research Foundation of South Africa (Grant Number: 129850). Computations were performed using facilities provided by the University of Cape Town's ICTS High Performance Computing team: hpc.uct.ac.za.",
+        "bbox": [
+            507,
+            250,
+            885,
+            347
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            510,
+            373,
+            608,
+            387
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "David Adelani, Md Mahfuz Ibn Alam, Antonios Anastasopoulos, Akshita Bhagia, Marta R. Costa-jussa, Jesse Dodge, Fahim Faisal, Christian Federmann, Natalia Fedorova, Francisco Guzmán, Sergey Koshelev, Jean Maillard, Vukosi Marivate, Jonathan Mbuya, Alexandre Mourachko, Safiyyah Saleem, Holger Schwenk, and Guillaume Wenzek. 2022. Findings of the WMT'22 shared task on large-scale machine translation evaluation for African languages. In Proceedings of the Seventh Conference on Machine Translation (WMT), pages 773-800, Abu Dhabi, United Arab Emirates (Hybrid). Association for Computational Linguistics.",
+            "Roee Aharoni, Melvin Johnson, and Orhan First. 2019. Massively multilingual neural machine translation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 3874-3884, Minneapolis, Minnesota. Association for Computational Linguistics.",
+            "Hyung Won Chung, Dan Garrette, Kiat Chuan Tan, and Jason Riesa. 2020. Improving multilingual models with language-clustered vocabularies. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4536-4546, Online. Association for Computational Linguistics.",
+            "Raj Dabre, Chenhui Chu, and Anoop Kunchukuttan. 2020. A survey of multilingual neural machine translation. ACM Comput. Surv., 53(5).",
+            "Orhan First, Kyunghyun Cho, and Yoshua Bengio. 2016. Multi-way, multilingual neural machine translation with a shared attention mechanism. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 866-875, San"
+        ],
+        "bbox": [
+            510,
+            395,
+            885,
+            921
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "2198",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Diego, California. Association for Computational Linguistics.",
+            "Naman Goyal, Cynthia Gao, Vishrav Chaudhary, Peng-Jen Chen, Guillaume Wenzek, Da Ju, Sanjana Krishnan, Marc'Aurelio Ranzato, Francisco Guzmán, and Angela Fan. 2021. The flores-101 evaluation benchmark for low-resource and multilingual machine translation.",
+            "Thanh-Le Ha, Jan Niehues, and Alex Waibel. 2016. Toward multilingual neural machine translation with universal encoder and decoder. In Proceedings of the 13th International Conference on Spoken Language Translation, Seattle, Washington D.C. International Workshop on Spoken Language Translation.",
+            "Taku Kudo. 2018. Subword regularization: Improving neural network translation models with multiple subword candidates. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 66-75, Melbourne, Australia. Association for Computational Linguistics.",
+            "Taku Kudo and John Richardson. 2018. SentencePiece: A simple and language independent subword tokenizer and tokenizer for neural text processing. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 66-71, Brussels, Belgium. Association for Computational Linguistics.",
+            "Antonis Maronikolakis, Philipp Dufter, and Hinrich Schütze. 2021. Wine is not v i n. on the compatibility of tokenizations across languages. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 2382-2399, Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Francois Meyer and Jan Buys. 2023. Subword segmental machine translation: Unifying segmentation and target sentence generation. In *Findings of the Association for Computational Linguistics: ACL* 2023, pages 2795-2809, Toronto, Canada. Association for Computational Linguistics.",
+            "Vaidehi Patil, Partha Talukdar, and Sunita Sarawagi. 2022. Overlap-based vocabulary generation improves cross-lingual transfer among related languages. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 219–233, Dublin, Ireland. Association for Computational Linguistics.",
+            "Telmo Pires, Eva Schlinger, and Dan Garrette. 2019. How multilingual is multilingual BERT? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4996-5001, Florence, Italy. Association for Computational Linguistics.",
+            "Maja Popovic. 2017.  $\\mathrm{chrF}++$ : words helping character n-grams. In Proceedings of the Second Conference on Machine Translation, pages 612-618, Copenhagen, Denmark. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            115,
+            85,
+            487,
+            919
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Rigardt Pretorius, Ansu Berg, Laurette Pretorius, and Biffie Viljoen. 2009. Setswana tokenisation and computational verb morphology: Facing the challenge of a disjunctive orthography. In Proceedings of the First Workshop on Language Technologies for African Languages, pages 66-73, Athens, Greece. Association for Computational Linguistics.",
+            "Phillip Rust, Jonas Pfeiffer, Ivan Vulic, Sebastian Ruder, and Iryna Gurevych. 2021. How good is your tokenizer? on the monolingual performance of multilingual language models. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 3118-3135, Online. Association for Computational Linguistics.",
+            "Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Neural machine translation of rare words with subword units. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1715-1725, Berlin, Germany. Association for Computational Linguistics.",
+            "Uri Shaham, Maha Elbayad, Vedanuj Goswami, Omer Levy, and Shruti Bhosale. 2023. Causes and cures for interference in multilingual translation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15849-15863, Toronto, Canada. Association for Computational Linguistics.",
+            "David Stap, Vlad Niculae, and Christof Monz. 2023. Viewing knowledge transfer in multilingual machine translation through a representational lens. In *Findings of the Association for Computational Linguistics: EMNLP* 2023, pages 14973-14987, Singapore. Association for Computational Linguistics.",
+            "Elsabé Taljard and Sonja E. Bosch. 2006. A comparison of approaches to word class tagging: Disjunctively vs. conjunctively written bantu languages. Nordic Journal of African Studies, 15:428-442.",
+            "NLLB Team, Marta R. Costa-jussà, James Cross, Onur Celebi, Maha Elbayad, Kenneth Heafield, Kevin Hef fernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, Anna Sun, Skyler Wang, Guillaume Wenzek, Al Youngblood, Bapi Akula, Loic Barrault, Gabriel Mejia Gonzalez, Prangthip Hansanti, John Hoffman, Semarley Jarrett, Kaushik Ram Sadagopan, Dirk Rowe, Shannon Spruit, Chau Tran, Pierre Andrews, Necip Fazil Ayan, Shruti Bhosale, Sergey Edunov, Angela Fan, Cynthia Gao, Vedanuj Goswami, Francisco Guzmán, Philipp Koehn, Alexandre Mourachko, Christophe Ropers, Safiyyah Saleem, Holger Schwenk, and Jeff Wang. 2022. No language left behind: Scaling humancentered machine translation.",
+            "Changhan Wang, Kyunghyun Cho, and Jiatao Gu. 2020a. Neural machine translation with byte-level subwords. In AAAI 2020 - 34th AAAI Conference on"
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            919
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "2199",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "ref_text",
+        "text": "Artificial Intelligence, AAAI 2020 - 34th AAAI Conference on Artificial Intelligence, pages 9154-9160. AAAI press. Publisher Copyright: Copyright © 2020 Association for the Advancement of Artificial Intelligence. All rights reserved.; 34th AAAI Conference on Artificial Intelligence, AAAI 2020; Conference date: 07-02-2020 Through 12-02-2020.",
+        "bbox": [
+            132,
+            85,
+            489,
+            178
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "ref_text",
+        "text": "Xinyi Wang, Sebastian Ruder, and Graham Neubig. 2021. Multi-view subword regularization. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 473-482, Online. Association for Computational Linguistics.",
+        "bbox": [
+            115,
+            187,
+            489,
+            280
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "ref_text",
+        "text": "Zihan Wang, Karthikeyan K, Stephen Mayhew, and Dan Roth. 2020b. Extending multilingual BERT to low-resource languages. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 2649-2656, Online. Association for Computational Linguistics.",
+        "bbox": [
+            115,
+            288,
+            489,
+            369
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "ref_text",
+        "text": "Shijie Wu and Mark Dredze. 2019. Beto, bentz, becas: The surprising cross-lingual effectiveness of BERT. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 833–844, Hong Kong, China. Association for Computational Linguistics.",
+        "bbox": [
+            115,
+            376,
+            489,
+            481
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "ref_text",
+        "text": "Judit Ács. 2019. Exploring BERT's vocabulary.",
+        "bbox": [
+            115,
+            491,
+            438,
+            507
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "A Model Configurations",
+        "text_level": 1,
+        "bbox": [
+            115,
+            519,
+            341,
+            535
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We use the model size and training hyperparameters of the fairseq transformer-base architecture. We train our models for 45 epochs and use a language sampling temperature of  $T = 1.5$  to balance exposure to low-resource and high-resource languages. In our cross-lingual finetuning experiments we finetune models on en→ss for 20 epochs and on en→xh/af/ts for 10 epochs.",
+        "bbox": [
+            115,
+            545,
+            489,
+            673
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "When applying subword methods we specify a shared vocabulary size of 8k. This is slightly smaller than the optimal vocabulary size of 10k used by Meyer and Buys (2023) in experiments on these same languages, since we use smaller subsets of the WMT22 datasets.",
+        "bbox": [
+            115,
+            674,
+            489,
+            768
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "2200",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 6
+    }
+]

2024/A Systematic Analysis of Subwords and Cross-Lingual Transfer in Multilingual Translation/d12825ff-73dc-4d7f-b285-a37a03d24322_model.json

@@ -0,0 +1,1314 @@
+[
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.233,
+                0.091,
+                0.766,
+                0.131
+            ],
+            "angle": 0,
+            "content": "A Systematic Analysis of Subwords and Cross-Linguual Transfer in Multilingual Translation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.365,
+                0.153,
+                0.638,
+                0.17
+            ],
+            "angle": 0,
+            "content": "Francois Meyer and Jan Buys"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.365,
+                0.17,
+                0.638,
+                0.186
+            ],
+            "angle": 0,
+            "content": "Department of Computer Science"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.399,
+                0.187,
+                0.604,
+                0.203
+            ],
+            "angle": 0,
+            "content": "University of Cape Town"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.28,
+                0.204,
+                0.724,
+                0.219
+            ],
+            "angle": 0,
+            "content": "francois.meyer@uct.ac.za, jbuys@cs.uct.ac.za"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.261,
+                0.261,
+                0.341,
+                0.277
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.145,
+                0.29,
+                0.461,
+                0.576
+            ],
+            "angle": 0,
+            "content": "Multilingual modelling can improve machine translation for low-resource languages, partly through shared subword representations. This paper studies the role of subword segmentation in cross-lingual transfer. We systematically compare the efficacy of several subword methods in promoting synergy and preventing interference across different linguistic typologies. Our findings show that subword regularisation boosts synergy in multilingual modelling, whereas BPE more effectively facilitates transfer during cross-lingual fine-tuning. Notably, our results suggest that differences in orthographic word boundary conventions (the morphological granularity of written words) may impede cross-lingual transfer more significantly than linguistic unrelatedness. Our study confirms that decisions around subword modelling can be key to optimising the benefits of multilingual modelling."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.589,
+                0.26,
+                0.604
+            ],
+            "angle": 0,
+            "content": "1 Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.615,
+                0.49,
+                0.807
+            ],
+            "angle": 0,
+            "content": "Machine translation (MT) models have become increasingly multilingual (Dabre et al., 2020). This greatly benefits low-resource languages through positive transfer from high-resource languages (Ha et al., 2016; Aharoni et al., 2019). However, increasing multilinguality in a limited shared parameter space can lead to suboptimal performance for high-resource languages (Firat et al., 2016; Team et al., 2022). There is a tradeoff between maximising positive cross-lingual transfer (also known as synergy) while minimising negative cross-lingual interaction (also known as interference)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.809,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Several modelling decisions affect synergy and interference in multilingual MT. Shaham et al. (2023) experimentally analysed the influence of factors like model size and language data proportions. One aspect their study failed to consider is subword segmentation. The shared subword vocabulary of multilingual models presents a similar"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.511,
+                0.261,
+                0.875,
+                0.426
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.509,
+                0.437,
+                0.883,
+                0.482
+            ],
+            "angle": 0,
+            "content": "Figure 1: Performance increase for English \\(\\rightarrow\\) Siswati through multilingual modelling varies greatly across subword methods and linguistic contexts."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.513,
+                0.885,
+                0.593
+            ],
+            "angle": 0,
+            "content": "trade-off as the shared parameter space - overlapping subword representations induce synergy, but having to represent multiple languages in a limited vocabulary can harm cross-lingual transfer (Chung et al., 2020; Rust et al., 2021; Patil et al., 2022)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.597,
+                0.885,
+                0.805
+            ],
+            "angle": 0,
+            "content": "In this paper we experimentally analyse the role of subword segmentation in multilingual and cross-lingual MT. Our goal is to compare different classes of subword methods with regards to their ability to induce synergy, reduce interference, and transfer knowledge during cross-lingual finetuning. We also investigate how cross-lingual transfer is influenced by the linguistic similarities of interacting languages, with particular focus on factors related to subword structure like morphological typology and orthographic word boundary conventions (the degree to which morphemes are concatenated or written as separate orthographic words)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.809,
+                0.887,
+                0.921
+            ],
+            "angle": 0,
+            "content": "We run experiments on translation from English to four linguistically diverse South African languages (see Table 1). This selection covers different levels of language relatedness, morphological complexity, and orthographic word granularity, allowing us to analyse how these factors interact with different subword methods to influence cross-"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.48,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "2194"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.22,
+                0.946,
+                0.779,
+                0.958
+            ],
+            "angle": 0,
+            "content": "Findings of the Association for Computational Linguistics: NAACL 2024, pages 2194-2200"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.293,
+                0.958,
+                0.705,
+                0.972
+            ],
+            "angle": 0,
+            "content": "June 16-21, 2024 ©2024 Association for Computational Linguistics"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.117,
+                0.082,
+                0.883,
+                0.167
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Language</td><td>Family</td><td>Morphology</td><td>Orthography</td><td>What is your name?</td><td>Thank you!</td></tr><tr><td>Siswati (ss)</td><td>NC/Bantu/Nguni</td><td>agglutinative</td><td>conjunctive</td><td>Ngubani ligama lakho?</td><td>Ngiyabonga!</td></tr><tr><td>isiXhosa (xh)</td><td>NC/Bantu/Nguni</td><td>agglutinative</td><td>conjunctive</td><td>Ungubani igama lakho?</td><td>Enkosi!</td></tr><tr><td>Setswana (ts)</td><td>NC/Bantu/Sotho-Tswana</td><td>agglutinative</td><td>disjunctive</td><td>Leina la gago ke mang?</td><td>Ke a leboga!</td></tr><tr><td>Afrikaans (af)</td><td>Indo-European/Germanic</td><td>analytic</td><td>disjunctive</td><td>Wat is jou naam?</td><td>Dankie!</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.12,
+                0.175,
+                0.877,
+                0.191
+            ],
+            "angle": 0,
+            "content": "Table 1: We vary the language modelled alongside Siswati to control relatedness, morphology, and orthography."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.216,
+                0.491,
+                0.328
+            ],
+            "angle": 0,
+            "content": "lingual transfer. Low-resource languages stand to benefit most from multilingual modelling. In all our experiments we focus on cross-lingual transfer to Siswati, which is by far the least resourced among the languages included. It presents exactly the type of real world low-resource translation scenario we are interested in studying."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.33,
+                0.49,
+                0.603
+            ],
+            "angle": 0,
+            "content": "We conduct two sets of experiments - multilingual MT and cross-lingual finetuning. Our multilingual experiments follow Shaham et al. (2023) in training several trilingual MT models and comparing synergy/interference (see Figure 1). In the cross-lingual finetuning experiments we finetune pretrained bilingual MT models on new languages. Our results demonstrate that decisions around subword segmentation significantly affect MT performance. ULM (Kudo, 2018) improves synergy in multilingual modelling, while BPE (Sennrich et al., 2016) enhances cross-lingual transfer during finetuning. Going beyond linguistic relatedness, we find that the much less studied influence of orthographic word boundary conventions can drastically affect the cross-lingual transfer achieved between interacting languages."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.619,
+                0.271,
+                0.634
+            ],
+            "angle": 0,
+            "content": "2 Related Work"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.647,
+                0.49,
+                0.856
+            ],
+            "angle": 0,
+            "content": "Synergy and interference are well-established phenomena (Firat et al., 2016; Aharoni et al., 2019; Team et al., 2022), but not well understood. Shatham et al. (2023) address this by systematically analysing the role of several factors in synergy and interference: (1) model size, (2) data size, (3) language proportions, (4) number of languages, and (5) language relatedness. Their results show that scaling model size and tuning the data sampling temperature greatly alleviates interference. They do not vary subword segmentation in their experiments, using the same sentencepiece (Kudo and Richardson, 2018) vocabulary across all models."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.858,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": "However, multilingual vocabularies are known to affect cross-lingual transfer through factors such as cross-lingual subword overlap (Pires et al., 2019; Wu and Dredze, 2019; Patil et al., 2022) and under"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.216,
+                0.885,
+                0.345
+            ],
+            "angle": 0,
+            "content": "represented low-resource languages (Wang et al., 2021; Acs, 2019). These issues have mainly been studied for multilingual language modelling (Rust et al., 2021; Maronikolakis et al., 2021; Chung et al., 2020), but the same concerns hold for MT (Wang et al., 2020a). We are unaware of existing work comparing the multitude of proposed subword methods in the context of multilingual MT."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.36,
+                0.659,
+                0.377
+            ],
+            "angle": 0,
+            "content": "3 Methodology"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.388,
+                0.885,
+                0.644
+            ],
+            "angle": 0,
+            "content": "This study involves two sets of MT experiments - (1) multilingual (trilingual) experiments to investigate synergy/interference, and (2) finetuning experiments to analyse cross-lingual transfer. Our goal is to determine which subwords benefit low-resource languages and how cross-lingual transfer depends on linguistic typology. The linguistic diversity of South Africa is an ideal testing ground for our purposes. Siswati is a low-resource agglutinative language, so effective subword modelling is critical for dealing with the inevitably high proportion of out-of-vocabulary words in held-out datasets. We use Siswati as the low-resource target language in our experiments and alternate the higher resourced language modelled alongside Siswati between isi-Xhosa, Setsswana, and Afrikaans."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.647,
+                0.886,
+                0.824
+            ],
+            "angle": 0,
+            "content": "Table 1 shows how these language present varying linguistic relationships to Siswati. IsiXhosa is closely related. Setswana is somewhat related and also agglutinative, but diverges in its orthography - its writing system is disjunctive (Pretorius et al., 2009). This refers to how linguistic words (e.g. nouns, verbs) are represented as orthographic words (space-separated tokens). Disjunctive orthographies write a single linguistic word as multiple orthographic words (e.g. in Setswana prefixal morphemes are space-separated from verbal roots)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.825,
+                0.885,
+                0.922
+            ],
+            "angle": 0,
+            "content": "While linguistic relatedness and morphological complexity are obvious features to consider in any analysis of cross-lingual interactions, we are unaware of work considering the impact of orthographic word boundary conventions. We include it as a factor in our study because of its potential"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "2195"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.112,
+                0.085,
+                0.49,
+                0.262
+            ],
+            "angle": 0,
+            "content": "relevance to subword segmentation. Orthographic word boundaries determine the pre-tokenisation of text before subword segmenters are applied, so it could well affect aspects like segmentation granularity and overlap between the subword vocabularies of different languages. Afrikaans is linguistically unrelated to Siswati and also disjunctive, but because of its analytic morphology (lower morpheme-to-word ratio) its written words are sometimes more aligned to those of Siswati (e.g. see phrases in Table 1)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.112,
+                0.265,
+                0.49,
+                0.41
+            ],
+            "angle": 0,
+            "content": "This selection of languages allows us to isolate the cross-lingual effects of linguistic relatedness, morphological typology, and orthographic word boundary conventions. In the case of Setswana-Siswati, we can study whether the potentially positive cross-lingual effect of their linguistic relatedness is nullified by the fact that the two languages have very different conventions for orthographic word boundaries."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.431,
+                0.345,
+                0.449
+            ],
+            "angle": 0,
+            "content": "3.1 Multilingual Modelling"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.459,
+                0.489,
+                0.492
+            ],
+            "angle": 0,
+            "content": "We train two bilingual models and one trilingual model per language pair (see Table 2)."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.118,
+                0.508,
+                0.486,
+                0.58
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Languages</td><td>Examples</td><td>Subwords</td></tr><tr><td>en→ss</td><td>166k</td><td>BPE/ULM/SSMT</td></tr><tr><td>en→xh/ts/af</td><td>1.6m</td><td>BPE/ULM/SSMT</td></tr><tr><td>en→ss+ xh/ts/af</td><td>1.6m+166k</td><td>BPE/ULM/SSMT/OBPE</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.113,
+                0.59,
+                0.487,
+                0.62
+            ],
+            "angle": 0,
+            "content": "Table 2: Multilingual experimental setup: bilingual and trilingual models (bilingual OBPE is equivalent to BPE)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.112,
+                0.645,
+                0.49,
+                0.789
+            ],
+            "angle": 0,
+            "content": "This setup allows us to compare how MT performance changes for en \\(\\rightarrow\\) ss and en \\(\\rightarrow\\) xh/ts/af going from bilingual models to multilingual models. Following Shaham et al. (2023), we measure synergy/interference for a translation direction \\(s\\to t\\) by the relative difference in performance between a bilingual model trained to translate only from \\(s\\) to \\(t\\) and a multilingual model trained to translate to an additional language."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.112,
+                0.793,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Shaham et al. (2023) use test set cross-entropy loss to measure MT performance, but this cannot be reliably used to compare across different subword segmentations. Instead, we use test set \\(\\mathrm{chrF}++\\) (Popovic, 2017) to measure performance. It is a reference-based metric that combines word and character information, so it is well suited for evaluating subword-level performance. Our modified"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.085,
+                0.854,
+                0.102
+            ],
+            "angle": 0,
+            "content": "formula for measuring synergy/interference is"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.52,
+                0.112,
+                0.871,
+                0.15
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {I} _ {s \\rightarrow t} = \\frac {\\mathrm {C H R F} + + (M _ {s \\rightarrow t} ^ {\\mathrm {m u l t i}}) - \\mathrm {C H R F} + + (M _ {s \\rightarrow t} ^ {\\mathrm {b i}})}{\\mathrm {C H R F} + + (M _ {s \\rightarrow t} ^ {\\mathrm {b i}})},\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.159,
+                0.883,
+                0.24
+            ],
+            "angle": 0,
+            "content": "where \\(M\\) are trained multilingual/bilingual models evaluated on \\(s\\rightarrow t\\) translation. Negative values of \\(I_{s\\rightarrow t}\\) indicate worse performance for \\(s\\rightarrow t\\) in the multilingual model (interference) and positive values indicate improved performance (synergy)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.251,
+                0.761,
+                0.267
+            ],
+            "angle": 0,
+            "content": "3.2 Cross-Lingual Finetuning"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.272,
+                0.885,
+                0.415
+            ],
+            "angle": 0,
+            "content": "We train a bilingual subword segmenter and MT model for en→xh/ts/af, and then finetune and evaluate the model in the other translation directions (e.g. pretrain en→xh and finetune on en→ss, en→ts, and en→af). Varying the subword method reveals how different subwords facilitate crosslingual transfer during finetuning from higher resourced languages (isiXhosa/Setswana/Afrikaans) to lower resourced Siswati."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.428,
+                0.719,
+                0.446
+            ],
+            "angle": 0,
+            "content": "4 Experimental Setup"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.454,
+                0.884,
+                0.55
+            ],
+            "angle": 0,
+            "content": "We compare five subword segmenters (four per experiment). We chose methods that represent the main paradigms of subword segmentation - deterministic segmentation, subword regularisation, learning subwords during training, and subword techniques for enhancing cross-lingual transfer."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.551,
+                0.884,
+                0.599
+            ],
+            "angle": 0,
+            "content": "1. BPE (Sennrich et al., 2016) iteratively adds frequently co-occurring subwords to its vocabulary. We use it as a deterministic segmenter."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.6,
+                0.885,
+                0.679
+            ],
+            "angle": 0,
+            "content": "2. ULM (Kudo, 2018) learns segmentation to optimise a unigram language model and can be used as a probabilistic segmenter, exposing models to multiple subword segmentations for regularisation. We set the sampling parameter \\(\\alpha\\) to 0.5."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.68,
+                0.886,
+                0.744
+            ],
+            "angle": 0,
+            "content": "3. SSMT (Meyer and Buys, 2023) is a subword segmental MT model which learns subword segmentation jointly during MT training, with the goal of learning subwords that optimise MT performance."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.745,
+                0.884,
+                0.825
+            ],
+            "angle": 0,
+            "content": "4. OBPE (Patil et al., 2022) modifies BPE to boost subword overlap among languages in multilingual vocabularies. We use it in our multilingual experiments to see if increased shared subword representations promote synergy."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.825,
+                0.884,
+                0.922
+            ],
+            "angle": 0,
+            "content": "5. XBPE (Wang et al., 2020b) extends the BPE vocabulary of a pretrained model to include BPE subwords of a new translation direction. New subword embeddings are randomly initialised. We use it in our finetuning experiments to see if the vocabulary extension enhances cross-lingual transfer."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.508,
+                0.551,
+                0.886,
+                0.922
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "2196"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.116,
+                0.083,
+                0.482,
+                0.212
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.114,
+                0.222,
+                0.489,
+                0.252
+            ],
+            "angle": 0,
+            "content": "Figure 2: Performance change for en→xh/af/ts through multilingual modelling alongside en→ss."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.124,
+                0.265,
+                0.481,
+                0.38
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Model</td><td>tgt</td><td>BPE</td><td>ULM</td><td>SSMT</td><td>OBPE</td></tr><tr><td rowspan=\"2\">en→ss/xh</td><td>ss</td><td>33.4</td><td>35.1</td><td>33.6</td><td>31.1</td></tr><tr><td>xh</td><td>46.8</td><td>47.2</td><td>46.1</td><td>44.6</td></tr><tr><td rowspan=\"2\">en→ss/af</td><td>ss</td><td>28.2</td><td>30.6</td><td>29.7</td><td>27.4</td></tr><tr><td>af</td><td>60.9</td><td>61.1</td><td>60.2</td><td>60.2</td></tr><tr><td rowspan=\"2\">en→ss/ts</td><td>ss</td><td>22.5</td><td>27.5</td><td>17.8</td><td>18.8</td></tr><tr><td>ts</td><td>31.3</td><td>34.7</td><td>23.6</td><td>26.4</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.145,
+                0.389,
+                0.454,
+                0.404
+            ],
+            "angle": 0,
+            "content": "Table 3: Test set \\( \\mathrm{{chrF}} +  + \\) of trilingual models."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.43,
+                0.49,
+                0.64
+            ],
+            "angle": 0,
+            "content": "Training We train models on WMT22 data (Adelani et al., 2022) and validate and test on FLORES (Goyal et al., 2021; Team et al., 2022). The number of training sentences are shown in Table 2. For en \\(\\rightarrow\\) ss this is the full WMT22 dataset, but for en \\(\\rightarrow\\) xh/ts/af we sampled sentences from en \\(\\rightarrow\\) xh and en \\(\\rightarrow\\) ts to match the size of en \\(\\rightarrow\\) af, removing data size as an influence. We also removed examples from en \\(\\rightarrow\\) xh/ts/af where English source sentences were found in en \\(\\rightarrow\\) ss to neutralise the positive transfer effect of multi-parallel overlap (Stap et al., 2023). The hyperparameters of our models and subword methods are included in Appendix A."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.652,
+                0.331,
+                0.669
+            ],
+            "angle": 0,
+            "content": "5 Results & Discussion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.678,
+                0.489,
+                0.76
+            ],
+            "angle": 0,
+            "content": "We plot the synergy/interference analysis of our multilingual experiments in Figures 1 & 2, while the absolute performance of the models are provided in Table 3. The results from our cross-lingual finetuning experiments are presented in Figure 3."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.772,
+                0.466,
+                0.803
+            ],
+            "angle": 0,
+            "content": "5.1 Which subwords promote synergy and minimise interference?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.809,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": "ULM consistently achieves greater synergy than other subword methods. This holds across all linguistic contexts (Fig. 1) and results in better absolute performance in all translation directions (Table 3). It comes at the cost of minimal interference for the higher resourced languages, and even some synergy for en→ts (Fig. 2). The subword regulari"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.51,
+                0.081,
+                0.875,
+                0.293
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.508,
+                0.302,
+                0.884,
+                0.348
+            ],
+            "angle": 0,
+            "content": "Figure 3: Test set \\(\\mathrm{chrF}++\\) of pretraining for en \\(\\rightarrow\\) xh/af/ts (rows) and finetuning on en \\(\\rightarrow\\) xh/af/ts/ss (columns). Diagonal entries are bilingual models with no finetuning."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.372,
+                0.882,
+                0.404
+            ],
+            "angle": 0,
+            "content": "sation of ULM ensures that models are more robust to the varied subwords of multilingual modelling."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.416,
+                0.882,
+                0.431
+            ],
+            "angle": 0,
+            "content": "5.2 Which subwords transfer cross-lingually?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.437,
+                0.884,
+                0.614
+            ],
+            "angle": 0,
+            "content": "BPE subwords exhibit the greatest cross-lingual transferability. In contrast to our multilingual findings, the subword regularisation of ULM proves a barrier to cross-lingual finetuning. ULM is a probabilistic segmenter that is sampled during training, but when the probabilistic model is based on one language and applied to another, its samples might yield highly inadequate subword units. The consistent deterministic segmentation of BPE allows the finetuned model to adapt to a new translation direction effectively."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.626,
+                0.858,
+                0.643
+            ],
+            "angle": 0,
+            "content": "5.3 What is the role of linguistic typology?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.648,
+                0.884,
+                0.888
+            ],
+            "angle": 0,
+            "content": "A consistent pattern emerges in the cross-lingual dynamics between Siswati and other languages. IsiXhosa modelling proves to be most beneficial for Siswati performance. Afrikaans achieves less transfer, presumably because it is not related. Somewhat surprisingly, the weakest synergy is between Siswati and Setswana, even though both are agglutinative Bantu languages. This highlights the impact of orthographic systems on cross-lingual transfer: diverging word boundary conventions can impede cross-lingual transfer more than linguistic unrelatedness. Data-driven multilingual models that learn from text might miss underlying similarities between languages that are obscured by superficial differences in their surface realisations."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.89,
+                0.885,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Our results highlight two interacting effects. Firstly, linguistic relatedness does play a role -"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "2197"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.493,
+                0.295
+            ],
+            "angle": 0,
+            "content": "isiXhosa consistently improves Siswati more than Setsswana and Afrikaans. Secondly, in the specific case of Setsswana-Siswati, their relatedness is rendered all but irrelevant by the fact that the languages have diverging orthographies. Afrikaans does not have the extremely disjunctive orthography of Setsswana so even though it is less related to Siswati than Setsswana, the orthography of Setsswana prevents transfer to Siswati. Linguistic distance plays a role in both cases (Afrikaans-Siswati and Setsswana-Siswati) but for Setsswana-Siswati it is a less important factor than orthography."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.296,
+                0.492,
+                0.489
+            ],
+            "angle": 0,
+            "content": "Orthography is a notable difference between Nguni languages like Siswati and Sotho-Tswana languages like Setsswana (Pretorius et al., 2009). Taljard and Bosch (2006) showed that the diverging orthographies of these two language groups necessitate different approaches to more traditional NLP tasks, even though the languages are linguistically and morphologically related. Our results suggest a similar situation for cross-lingual transfer in multilingual modelling: Differences in orthographic word boundary conventions harms synergy between otherwise related languages."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.501,
+                0.248,
+                0.516
+            ],
+            "angle": 0,
+            "content": "6 Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.528,
+                0.49,
+                0.786
+            ],
+            "angle": 0,
+            "content": "We presented an in-depth study on the role of subwords in multilingual and cross-lingual MT. Our results demonstrate that subword segmentation significantly influences cross-lingual interactions. ULM proves optimal for transfer to low-resource languages in multilingual modelling, while BPE enables greater cross-lingual transfer during finetuning. Besides language relatedness, we show that similarities/differences in orthographic word granularity can greatly affect multilingual performance. There is more work to be done on the role of orthographic word boundary conventions in neural MT. Future work could aim to design multilingual techniques that see past orthographic differences in order to leverage more fundamental similarities between languages."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.799,
+                0.221,
+                0.814
+            ],
+            "angle": 0,
+            "content": "Limitations"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.825,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Our study is limited to translation from English to four South African languages. While the chosen languages are typologically diverse, our conclusions might not necessarily hold for languages from different language families and with distinct orthographies. We did not consider languages that"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.887,
+                0.214
+            ],
+            "angle": 0,
+            "content": "have multiple orthographies, which might be another approach to study the effects of orthography. The performance differences between different subword segmentation methods across languages in our results are relatively consistent, but a more detailed analysis on the interaction between choice of subword segmentation method and language could yield additional explanations of the results."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.226,
+                0.683,
+                0.241
+            ],
+            "angle": 0,
+            "content": "Acknowledgements"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.251,
+                0.886,
+                0.348
+            ],
+            "angle": 0,
+            "content": "This work is based on research supported in part by the National Research Foundation of South Africa (Grant Number: 129850). Computations were performed using facilities provided by the University of Cape Town's ICTS High Performance Computing team: hpc.uct.ac.za."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.511,
+                0.374,
+                0.61,
+                0.388
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.396,
+                0.886,
+                0.568
+            ],
+            "angle": 0,
+            "content": "David Adelani, Md Mahfuz Ibn Alam, Antonios Anastasopoulos, Akshita Bhagia, Marta R. Costa-jussa, Jesse Dodge, Fahim Faisal, Christian Federmann, Natalia Fedorova, Francisco Guzmán, Sergey Koshelev, Jean Maillard, Vukosi Marivate, Jonathan Mbuya, Alexandre Mourachko, Safiyyah Saleem, Holger Schwenk, and Guillaume Wenzek. 2022. Findings of the WMT'22 shared task on large-scale machine translation evaluation for African languages. In Proceedings of the Seventh Conference on Machine Translation (WMT), pages 773-800, Abu Dhabi, United Arab Emirates (Hybrid). Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.576,
+                0.887,
+                0.682
+            ],
+            "angle": 0,
+            "content": "Roee Aharoni, Melvin Johnson, and Orhan First. 2019. Massively multilingual neural machine translation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 3874-3884, Minneapolis, Minnesota. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.69,
+                0.886,
+                0.784
+            ],
+            "angle": 0,
+            "content": "Hyung Won Chung, Dan Garrette, Kiat Chuan Tan, and Jason Riesa. 2020. Improving multilingual models with language-clustered vocabularies. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4536-4546, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.792,
+                0.886,
+                0.833
+            ],
+            "angle": 0,
+            "content": "Raj Dabre, Chenhui Chu, and Anoop Kunchukuttan. 2020. A survey of multilingual neural machine translation. ACM Comput. Surv., 53(5)."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.842,
+                0.886,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Orhan First, Kyunghyun Cho, and Yoshua Bengio. 2016. Multi-way, multilingual neural machine translation with a shared attention mechanism. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 866-875, San"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.511,
+                0.396,
+                0.887,
+                0.922
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "2198"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.135,
+                0.086,
+                0.486,
+                0.113
+            ],
+            "angle": 0,
+            "content": "Diego, California. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.121,
+                0.488,
+                0.199
+            ],
+            "angle": 0,
+            "content": "Naman Goyal, Cynthia Gao, Vishrav Chaudhary, Peng-Jen Chen, Guillaume Wenzek, Da Ju, Sanjana Krishnan, Marc'Aurelio Ranzato, Francisco Guzmán, and Angela Fan. 2021. The flores-101 evaluation benchmark for low-resource and multilingual machine translation."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.208,
+                0.487,
+                0.287
+            ],
+            "angle": 0,
+            "content": "Thanh-Le Ha, Jan Niehues, and Alex Waibel. 2016. Toward multilingual neural machine translation with universal encoder and decoder. In Proceedings of the 13th International Conference on Spoken Language Translation, Seattle, Washington D.C. International Workshop on Spoken Language Translation."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.295,
+                0.487,
+                0.386
+            ],
+            "angle": 0,
+            "content": "Taku Kudo. 2018. Subword regularization: Improving neural network translation models with multiple subword candidates. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 66-75, Melbourne, Australia. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.395,
+                0.487,
+                0.486
+            ],
+            "angle": 0,
+            "content": "Taku Kudo and John Richardson. 2018. SentencePiece: A simple and language independent subword tokenizer and tokenizer for neural text processing. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 66-71, Brussels, Belgium. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.495,
+                0.487,
+                0.573
+            ],
+            "angle": 0,
+            "content": "Antonis Maronikolakis, Philipp Dufter, and Hinrich Schütze. 2021. Wine is not v i n. on the compatibility of tokenizations across languages. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 2382-2399, Punta Cana, Dominican Republic. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.582,
+                0.487,
+                0.66
+            ],
+            "angle": 0,
+            "content": "Francois Meyer and Jan Buys. 2023. Subword segmental machine translation: Unifying segmentation and target sentence generation. In *Findings of the Association for Computational Linguistics: ACL* 2023, pages 2795-2809, Toronto, Canada. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.669,
+                0.487,
+                0.759
+            ],
+            "angle": 0,
+            "content": "Vaidehi Patil, Partha Talukdar, and Sunita Sarawagi. 2022. Overlap-based vocabulary generation improves cross-lingual transfer among related languages. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 219–233, Dublin, Ireland. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.768,
+                0.487,
+                0.845
+            ],
+            "angle": 0,
+            "content": "Telmo Pires, Eva Schlinger, and Dan Garrette. 2019. How multilingual is multilingual BERT? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4996-5001, Florence, Italy. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.855,
+                0.487,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Maja Popovic. 2017. \\(\\mathrm{chrF}++\\): words helping character n-grams. In Proceedings of the Second Conference on Machine Translation, pages 612-618, Copenhagen, Denmark. Association for Computational Linguistics."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.086,
+                0.488,
+                0.92
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.086,
+                0.882,
+                0.178
+            ],
+            "angle": 0,
+            "content": "Rigardt Pretorius, Ansu Berg, Laurette Pretorius, and Biffie Viljoen. 2009. Setswana tokenisation and computational verb morphology: Facing the challenge of a disjunctive orthography. In Proceedings of the First Workshop on Language Technologies for African Languages, pages 66-73, Athens, Greece. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.189,
+                0.882,
+                0.306
+            ],
+            "angle": 0,
+            "content": "Phillip Rust, Jonas Pfeiffer, Ivan Vulic, Sebastian Ruder, and Iryna Gurevych. 2021. How good is your tokenizer? on the monolingual performance of multilingual language models. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 3118-3135, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.317,
+                0.882,
+                0.409
+            ],
+            "angle": 0,
+            "content": "Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Neural machine translation of rare words with subword units. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1715-1725, Berlin, Germany. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.42,
+                0.882,
+                0.511
+            ],
+            "angle": 0,
+            "content": "Uri Shaham, Maha Elbayad, Vedanuj Goswami, Omer Levy, and Shruti Bhosale. 2023. Causes and cures for interference in multilingual translation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15849-15863, Toronto, Canada. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.522,
+                0.882,
+                0.6
+            ],
+            "angle": 0,
+            "content": "David Stap, Vlad Niculae, and Christof Monz. 2023. Viewing knowledge transfer in multilingual machine translation through a representational lens. In *Findings of the Association for Computational Linguistics: EMNLP* 2023, pages 14973-14987, Singapore. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.611,
+                0.882,
+                0.663
+            ],
+            "angle": 0,
+            "content": "Elsabé Taljard and Sonja E. Bosch. 2006. A comparison of approaches to word class tagging: Disjunctively vs. conjunctively written bantu languages. Nordic Journal of African Studies, 15:428-442."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.675,
+                0.882,
+                0.87
+            ],
+            "angle": 0,
+            "content": "NLLB Team, Marta R. Costa-jussà, James Cross, Onur Celebi, Maha Elbayad, Kenneth Heafield, Kevin Hef fernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, Anna Sun, Skyler Wang, Guillaume Wenzek, Al Youngblood, Bapi Akula, Loic Barrault, Gabriel Mejia Gonzalez, Prangthip Hansanti, John Hoffman, Semarley Jarrett, Kaushik Ram Sadagopan, Dirk Rowe, Shannon Spruit, Chau Tran, Pierre Andrews, Necip Fazil Ayan, Shruti Bhosale, Sergey Edunov, Angela Fan, Cynthia Gao, Vedanuj Goswami, Francisco Guzmán, Philipp Koehn, Alexandre Mourachko, Christophe Ropers, Safiyyah Saleem, Holger Schwenk, and Jeff Wang. 2022. No language left behind: Scaling humancentered machine translation."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.881,
+                0.882,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Changhan Wang, Kyunghyun Cho, and Jiatao Gu. 2020a. Neural machine translation with byte-level subwords. In AAAI 2020 - 34th AAAI Conference on"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.512,
+                0.086,
+                0.882,
+                0.92
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "2199"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.133,
+                0.086,
+                0.49,
+                0.179
+            ],
+            "angle": 0,
+            "content": "Artificial Intelligence, AAAI 2020 - 34th AAAI Conference on Artificial Intelligence, pages 9154-9160. AAAI press. Publisher Copyright: Copyright © 2020 Association for the Advancement of Artificial Intelligence. All rights reserved.; 34th AAAI Conference on Artificial Intelligence, AAAI 2020; Conference date: 07-02-2020 Through 12-02-2020."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.188,
+                0.49,
+                0.281
+            ],
+            "angle": 0,
+            "content": "Xinyi Wang, Sebastian Ruder, and Graham Neubig. 2021. Multi-view subword regularization. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 473-482, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.29,
+                0.49,
+                0.37
+            ],
+            "angle": 0,
+            "content": "Zihan Wang, Karthikeyan K, Stephen Mayhew, and Dan Roth. 2020b. Extending multilingual BERT to low-resource languages. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 2649-2656, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.378,
+                0.49,
+                0.482
+            ],
+            "angle": 0,
+            "content": "Shijie Wu and Mark Dredze. 2019. Beto, bentz, becas: The surprising cross-lingual effectiveness of BERT. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 833–844, Hong Kong, China. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.492,
+                0.44,
+                0.508
+            ],
+            "angle": 0,
+            "content": "Judit Ács. 2019. Exploring BERT's vocabulary."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.117,
+                0.52,
+                0.342,
+                0.536
+            ],
+            "angle": 0,
+            "content": "A Model Configurations"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.116,
+                0.546,
+                0.49,
+                0.674
+            ],
+            "angle": 0,
+            "content": "We use the model size and training hyperparameters of the fairseq transformer-base architecture. We train our models for 45 epochs and use a language sampling temperature of \\( T = 1.5 \\) to balance exposure to low-resource and high-resource languages. In our cross-lingual finetuning experiments we finetune models on en→ss for 20 epochs and on en→xh/af/ts for 10 epochs."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.116,
+                0.675,
+                0.49,
+                0.769
+            ],
+            "angle": 0,
+            "content": "When applying subword methods we specify a shared vocabulary size of 8k. This is slightly smaller than the optimal vocabulary size of 10k used by Meyer and Buys (2023) in experiments on these same languages, since we use smaller subsets of the WMT22 datasets."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "2200"
+        }
+    ]
+]

2024/A Systematic Analysis of Subwords and Cross-Lingual Transfer in Multilingual Translation/d12825ff-73dc-4d7f-b285-a37a03d24322_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:df94fafbf0cae76727c96c7486b8653b184be3700a0736c36f29524a148459a3
+size 238167

2024/A Systematic Analysis of Subwords and Cross-Lingual Transfer in Multilingual Translation/full.md

@@ -0,0 +1,184 @@
+# A Systematic Analysis of Subwords and Cross-Linguual Transfer in Multilingual Translation
+
+Francois Meyer and Jan Buys
+
+Department of Computer Science
+
+University of Cape Town
+
+francois.meyer@uct.ac.za, jbuys@cs.uct.ac.za
+
+# Abstract
+
+Multilingual modelling can improve machine translation for low-resource languages, partly through shared subword representations. This paper studies the role of subword segmentation in cross-lingual transfer. We systematically compare the efficacy of several subword methods in promoting synergy and preventing interference across different linguistic typologies. Our findings show that subword regularisation boosts synergy in multilingual modelling, whereas BPE more effectively facilitates transfer during cross-lingual fine-tuning. Notably, our results suggest that differences in orthographic word boundary conventions (the morphological granularity of written words) may impede cross-lingual transfer more significantly than linguistic unrelatedness. Our study confirms that decisions around subword modelling can be key to optimising the benefits of multilingual modelling.
+
+# 1 Introduction
+
+Machine translation (MT) models have become increasingly multilingual (Dabre et al., 2020). This greatly benefits low-resource languages through positive transfer from high-resource languages (Ha et al., 2016; Aharoni et al., 2019). However, increasing multilinguality in a limited shared parameter space can lead to suboptimal performance for high-resource languages (Firat et al., 2016; Team et al., 2022). There is a tradeoff between maximising positive cross-lingual transfer (also known as synergy) while minimising negative cross-lingual interaction (also known as interference).
+
+Several modelling decisions affect synergy and interference in multilingual MT. Shaham et al. (2023) experimentally analysed the influence of factors like model size and language data proportions. One aspect their study failed to consider is subword segmentation. The shared subword vocabulary of multilingual models presents a similar
+
+![](images/c3659bb69add3615263b79c8fbafb932be5dff55dabf9eaa7f9442b17ae84bea.jpg)  
+Figure 1: Performance increase for English  $\rightarrow$  Siswati through multilingual modelling varies greatly across subword methods and linguistic contexts.
+
+trade-off as the shared parameter space - overlapping subword representations induce synergy, but having to represent multiple languages in a limited vocabulary can harm cross-lingual transfer (Chung et al., 2020; Rust et al., 2021; Patil et al., 2022).
+
+In this paper we experimentally analyse the role of subword segmentation in multilingual and cross-lingual MT. Our goal is to compare different classes of subword methods with regards to their ability to induce synergy, reduce interference, and transfer knowledge during cross-lingual finetuning. We also investigate how cross-lingual transfer is influenced by the linguistic similarities of interacting languages, with particular focus on factors related to subword structure like morphological typology and orthographic word boundary conventions (the degree to which morphemes are concatenated or written as separate orthographic words).
+
+We run experiments on translation from English to four linguistically diverse South African languages (see Table 1). This selection covers different levels of language relatedness, morphological complexity, and orthographic word granularity, allowing us to analyse how these factors interact with different subword methods to influence cross-
+
+<table><tr><td>Language</td><td>Family</td><td>Morphology</td><td>Orthography</td><td>What is your name?</td><td>Thank you!</td></tr><tr><td>Siswati (ss)</td><td>NC/Bantu/Nguni</td><td>agglutinative</td><td>conjunctive</td><td>Ngubani ligama lakho?</td><td>Ngiyabonga!</td></tr><tr><td>isiXhosa (xh)</td><td>NC/Bantu/Nguni</td><td>agglutinative</td><td>conjunctive</td><td>Ungubani igama lakho?</td><td>Enkosi!</td></tr><tr><td>Setswana (ts)</td><td>NC/Bantu/Sotho-Tswana</td><td>agglutinative</td><td>disjunctive</td><td>Leina la gago ke mang?</td><td>Ke a leboga!</td></tr><tr><td>Afrikaans (af)</td><td>Indo-European/Germanic</td><td>analytic</td><td>disjunctive</td><td>Wat is jou naam?</td><td>Dankie!</td></tr></table>
+
+Table 1: We vary the language modelled alongside Siswati to control relatedness, morphology, and orthography.
+
+lingual transfer. Low-resource languages stand to benefit most from multilingual modelling. In all our experiments we focus on cross-lingual transfer to Siswati, which is by far the least resourced among the languages included. It presents exactly the type of real world low-resource translation scenario we are interested in studying.
+
+We conduct two sets of experiments - multilingual MT and cross-lingual finetuning. Our multilingual experiments follow Shaham et al. (2023) in training several trilingual MT models and comparing synergy/interference (see Figure 1). In the cross-lingual finetuning experiments we finetune pretrained bilingual MT models on new languages. Our results demonstrate that decisions around subword segmentation significantly affect MT performance. ULM (Kudo, 2018) improves synergy in multilingual modelling, while BPE (Sennrich et al., 2016) enhances cross-lingual transfer during finetuning. Going beyond linguistic relatedness, we find that the much less studied influence of orthographic word boundary conventions can drastically affect the cross-lingual transfer achieved between interacting languages.
+
+# 2 Related Work
+
+Synergy and interference are well-established phenomena (Firat et al., 2016; Aharoni et al., 2019; Team et al., 2022), but not well understood. Shatham et al. (2023) address this by systematically analysing the role of several factors in synergy and interference: (1) model size, (2) data size, (3) language proportions, (4) number of languages, and (5) language relatedness. Their results show that scaling model size and tuning the data sampling temperature greatly alleviates interference. They do not vary subword segmentation in their experiments, using the same sentencepiece (Kudo and Richardson, 2018) vocabulary across all models.
+
+However, multilingual vocabularies are known to affect cross-lingual transfer through factors such as cross-lingual subword overlap (Pires et al., 2019; Wu and Dredze, 2019; Patil et al., 2022) and under
+
+represented low-resource languages (Wang et al., 2021; Acs, 2019). These issues have mainly been studied for multilingual language modelling (Rust et al., 2021; Maronikolakis et al., 2021; Chung et al., 2020), but the same concerns hold for MT (Wang et al., 2020a). We are unaware of existing work comparing the multitude of proposed subword methods in the context of multilingual MT.
+
+# 3 Methodology
+
+This study involves two sets of MT experiments - (1) multilingual (trilingual) experiments to investigate synergy/interference, and (2) finetuning experiments to analyse cross-lingual transfer. Our goal is to determine which subwords benefit low-resource languages and how cross-lingual transfer depends on linguistic typology. The linguistic diversity of South Africa is an ideal testing ground for our purposes. Siswati is a low-resource agglutinative language, so effective subword modelling is critical for dealing with the inevitably high proportion of out-of-vocabulary words in held-out datasets. We use Siswati as the low-resource target language in our experiments and alternate the higher resourced language modelled alongside Siswati between isi-Xhosa, Setsswana, and Afrikaans.
+
+Table 1 shows how these language present varying linguistic relationships to Siswati. IsiXhosa is closely related. Setswana is somewhat related and also agglutinative, but diverges in its orthography - its writing system is disjunctive (Pretorius et al., 2009). This refers to how linguistic words (e.g. nouns, verbs) are represented as orthographic words (space-separated tokens). Disjunctive orthographies write a single linguistic word as multiple orthographic words (e.g. in Setswana prefixal morphemes are space-separated from verbal roots).
+
+While linguistic relatedness and morphological complexity are obvious features to consider in any analysis of cross-lingual interactions, we are unaware of work considering the impact of orthographic word boundary conventions. We include it as a factor in our study because of its potential
+
+relevance to subword segmentation. Orthographic word boundaries determine the pre-tokenisation of text before subword segmenters are applied, so it could well affect aspects like segmentation granularity and overlap between the subword vocabularies of different languages. Afrikaans is linguistically unrelated to Siswati and also disjunctive, but because of its analytic morphology (lower morpheme-to-word ratio) its written words are sometimes more aligned to those of Siswati (e.g. see phrases in Table 1).
+
+This selection of languages allows us to isolate the cross-lingual effects of linguistic relatedness, morphological typology, and orthographic word boundary conventions. In the case of Setswana-Siswati, we can study whether the potentially positive cross-lingual effect of their linguistic relatedness is nullified by the fact that the two languages have very different conventions for orthographic word boundaries.
+
+# 3.1 Multilingual Modelling
+
+We train two bilingual models and one trilingual model per language pair (see Table 2).
+
+<table><tr><td>Languages</td><td>Examples</td><td>Subwords</td></tr><tr><td>en→ss</td><td>166k</td><td>BPE/ULM/SSMT</td></tr><tr><td>en→xh/ts/af</td><td>1.6m</td><td>BPE/ULM/SSMT</td></tr><tr><td>en→ss+ xh/ts/af</td><td>1.6m+166k</td><td>BPE/ULM/SSMT/OBPE</td></tr></table>
+
+Table 2: Multilingual experimental setup: bilingual and trilingual models (bilingual OBPE is equivalent to BPE).
+
+This setup allows us to compare how MT performance changes for en  $\rightarrow$  ss and en  $\rightarrow$  xh/ts/af going from bilingual models to multilingual models. Following Shaham et al. (2023), we measure synergy/interference for a translation direction  $s\to t$  by the relative difference in performance between a bilingual model trained to translate only from  $s$  to  $t$  and a multilingual model trained to translate to an additional language.
+
+Shaham et al. (2023) use test set cross-entropy loss to measure MT performance, but this cannot be reliably used to compare across different subword segmentations. Instead, we use test set  $\mathrm{chrF}++$  (Popovic, 2017) to measure performance. It is a reference-based metric that combines word and character information, so it is well suited for evaluating subword-level performance. Our modified
+
+formula for measuring synergy/interference is
+
+$$
+\mathcal {I} _ {s \rightarrow t} = \frac {\mathrm {C H R F} + + (M _ {s \rightarrow t} ^ {\mathrm {m u l t i}}) - \mathrm {C H R F} + + (M _ {s \rightarrow t} ^ {\mathrm {b i}})}{\mathrm {C H R F} + + (M _ {s \rightarrow t} ^ {\mathrm {b i}})},
+$$
+
+where  $M$  are trained multilingual/bilingual models evaluated on  $s\rightarrow t$  translation. Negative values of  $I_{s\rightarrow t}$  indicate worse performance for  $s\rightarrow t$  in the multilingual model (interference) and positive values indicate improved performance (synergy).
+
+# 3.2 Cross-Lingual Finetuning
+
+We train a bilingual subword segmenter and MT model for en→xh/ts/af, and then finetune and evaluate the model in the other translation directions (e.g. pretrain en→xh and finetune on en→ss, en→ts, and en→af). Varying the subword method reveals how different subwords facilitate crosslingual transfer during finetuning from higher resourced languages (isiXhosa/Setswana/Afrikaans) to lower resourced Siswati.
+
+# 4 Experimental Setup
+
+We compare five subword segmenters (four per experiment). We chose methods that represent the main paradigms of subword segmentation - deterministic segmentation, subword regularisation, learning subwords during training, and subword techniques for enhancing cross-lingual transfer.
+
+1. BPE (Sennrich et al., 2016) iteratively adds frequently co-occurring subwords to its vocabulary. We use it as a deterministic segmenter.  
+2. ULM (Kudo, 2018) learns segmentation to optimise a unigram language model and can be used as a probabilistic segmenter, exposing models to multiple subword segmentations for regularisation. We set the sampling parameter  $\alpha$  to 0.5.  
+3. SSMT (Meyer and Buys, 2023) is a subword segmental MT model which learns subword segmentation jointly during MT training, with the goal of learning subwords that optimise MT performance.  
+4. OBPE (Patil et al., 2022) modifies BPE to boost subword overlap among languages in multilingual vocabularies. We use it in our multilingual experiments to see if increased shared subword representations promote synergy.  
+5. XBPE (Wang et al., 2020b) extends the BPE vocabulary of a pretrained model to include BPE subwords of a new translation direction. New subword embeddings are randomly initialised. We use it in our finetuning experiments to see if the vocabulary extension enhances cross-lingual transfer.
+
+![](images/fe341bc27bd12b8ea9a8921aaf6c92d34d1f5543463deb49e678a0af795d87b7.jpg)  
+Figure 2: Performance change for en→xh/af/ts through multilingual modelling alongside en→ss.
+
+<table><tr><td>Model</td><td>tgt</td><td>BPE</td><td>ULM</td><td>SSMT</td><td>OBPE</td></tr><tr><td rowspan="2">en→ss/xh</td><td>ss</td><td>33.4</td><td>35.1</td><td>33.6</td><td>31.1</td></tr><tr><td>xh</td><td>46.8</td><td>47.2</td><td>46.1</td><td>44.6</td></tr><tr><td rowspan="2">en→ss/af</td><td>ss</td><td>28.2</td><td>30.6</td><td>29.7</td><td>27.4</td></tr><tr><td>af</td><td>60.9</td><td>61.1</td><td>60.2</td><td>60.2</td></tr><tr><td rowspan="2">en→ss/ts</td><td>ss</td><td>22.5</td><td>27.5</td><td>17.8</td><td>18.8</td></tr><tr><td>ts</td><td>31.3</td><td>34.7</td><td>23.6</td><td>26.4</td></tr></table>
+
+Table 3: Test set  $\mathrm{{chrF}} +  +$  of trilingual models.
+
+Training We train models on WMT22 data (Adelani et al., 2022) and validate and test on FLORES (Goyal et al., 2021; Team et al., 2022). The number of training sentences are shown in Table 2. For en  $\rightarrow$  ss this is the full WMT22 dataset, but for en  $\rightarrow$  xh/ts/af we sampled sentences from en  $\rightarrow$  xh and en  $\rightarrow$  ts to match the size of en  $\rightarrow$  af, removing data size as an influence. We also removed examples from en  $\rightarrow$  xh/ts/af where English source sentences were found in en  $\rightarrow$  ss to neutralise the positive transfer effect of multi-parallel overlap (Stap et al., 2023). The hyperparameters of our models and subword methods are included in Appendix A.
+
+# 5 Results & Discussion
+
+We plot the synergy/interference analysis of our multilingual experiments in Figures 1 & 2, while the absolute performance of the models are provided in Table 3. The results from our cross-lingual finetuning experiments are presented in Figure 3.
+
+# 5.1 Which subwords promote synergy and minimise interference?
+
+ULM consistently achieves greater synergy than other subword methods. This holds across all linguistic contexts (Fig. 1) and results in better absolute performance in all translation directions (Table 3). It comes at the cost of minimal interference for the higher resourced languages, and even some synergy for en→ts (Fig. 2). The subword regulari
+
+![](images/896c6c07ad2f3f3d7b46abd1f15f9fc6aa762874dbf3403014fd0282d0267c29.jpg)  
+Figure 3: Test set  $\mathrm{chrF}++$  of pretraining for en  $\rightarrow$  xh/af/ts (rows) and finetuning on en  $\rightarrow$  xh/af/ts/ss (columns). Diagonal entries are bilingual models with no finetuning.
+
+sation of ULM ensures that models are more robust to the varied subwords of multilingual modelling.
+
+# 5.2 Which subwords transfer cross-lingually?
+
+BPE subwords exhibit the greatest cross-lingual transferability. In contrast to our multilingual findings, the subword regularisation of ULM proves a barrier to cross-lingual finetuning. ULM is a probabilistic segmenter that is sampled during training, but when the probabilistic model is based on one language and applied to another, its samples might yield highly inadequate subword units. The consistent deterministic segmentation of BPE allows the finetuned model to adapt to a new translation direction effectively.
+
+# 5.3 What is the role of linguistic typology?
+
+A consistent pattern emerges in the cross-lingual dynamics between Siswati and other languages. IsiXhosa modelling proves to be most beneficial for Siswati performance. Afrikaans achieves less transfer, presumably because it is not related. Somewhat surprisingly, the weakest synergy is between Siswati and Setswana, even though both are agglutinative Bantu languages. This highlights the impact of orthographic systems on cross-lingual transfer: diverging word boundary conventions can impede cross-lingual transfer more than linguistic unrelatedness. Data-driven multilingual models that learn from text might miss underlying similarities between languages that are obscured by superficial differences in their surface realisations.
+
+Our results highlight two interacting effects. Firstly, linguistic relatedness does play a role -
+
+isiXhosa consistently improves Siswati more than Setsswana and Afrikaans. Secondly, in the specific case of Setsswana-Siswati, their relatedness is rendered all but irrelevant by the fact that the languages have diverging orthographies. Afrikaans does not have the extremely disjunctive orthography of Setsswana so even though it is less related to Siswati than Setsswana, the orthography of Setsswana prevents transfer to Siswati. Linguistic distance plays a role in both cases (Afrikaans-Siswati and Setsswana-Siswati) but for Setsswana-Siswati it is a less important factor than orthography.
+
+Orthography is a notable difference between Nguni languages like Siswati and Sotho-Tswana languages like Setsswana (Pretorius et al., 2009). Taljard and Bosch (2006) showed that the diverging orthographies of these two language groups necessitate different approaches to more traditional NLP tasks, even though the languages are linguistically and morphologically related. Our results suggest a similar situation for cross-lingual transfer in multilingual modelling: Differences in orthographic word boundary conventions harms synergy between otherwise related languages.
+
+# 6 Conclusion
+
+We presented an in-depth study on the role of subwords in multilingual and cross-lingual MT. Our results demonstrate that subword segmentation significantly influences cross-lingual interactions. ULM proves optimal for transfer to low-resource languages in multilingual modelling, while BPE enables greater cross-lingual transfer during finetuning. Besides language relatedness, we show that similarities/differences in orthographic word granularity can greatly affect multilingual performance. There is more work to be done on the role of orthographic word boundary conventions in neural MT. Future work could aim to design multilingual techniques that see past orthographic differences in order to leverage more fundamental similarities between languages.
+
+# Limitations
+
+Our study is limited to translation from English to four South African languages. While the chosen languages are typologically diverse, our conclusions might not necessarily hold for languages from different language families and with distinct orthographies. We did not consider languages that
+
+have multiple orthographies, which might be another approach to study the effects of orthography. The performance differences between different subword segmentation methods across languages in our results are relatively consistent, but a more detailed analysis on the interaction between choice of subword segmentation method and language could yield additional explanations of the results.
+
+# Acknowledgements
+
+This work is based on research supported in part by the National Research Foundation of South Africa (Grant Number: 129850). Computations were performed using facilities provided by the University of Cape Town's ICTS High Performance Computing team: hpc.uct.ac.za.
+
+# References
+
+David Adelani, Md Mahfuz Ibn Alam, Antonios Anastasopoulos, Akshita Bhagia, Marta R. Costa-jussa, Jesse Dodge, Fahim Faisal, Christian Federmann, Natalia Fedorova, Francisco Guzmán, Sergey Koshelev, Jean Maillard, Vukosi Marivate, Jonathan Mbuya, Alexandre Mourachko, Safiyyah Saleem, Holger Schwenk, and Guillaume Wenzek. 2022. Findings of the WMT'22 shared task on large-scale machine translation evaluation for African languages. In Proceedings of the Seventh Conference on Machine Translation (WMT), pages 773-800, Abu Dhabi, United Arab Emirates (Hybrid). Association for Computational Linguistics.  
+Roee Aharoni, Melvin Johnson, and Orhan First. 2019. Massively multilingual neural machine translation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 3874-3884, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Hyung Won Chung, Dan Garrette, Kiat Chuan Tan, and Jason Riesa. 2020. Improving multilingual models with language-clustered vocabularies. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4536-4546, Online. Association for Computational Linguistics.  
+Raj Dabre, Chenhui Chu, and Anoop Kunchukuttan. 2020. A survey of multilingual neural machine translation. ACM Comput. Surv., 53(5).  
+Orhan First, Kyunghyun Cho, and Yoshua Bengio. 2016. Multi-way, multilingual neural machine translation with a shared attention mechanism. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 866-875, San
+
+Diego, California. Association for Computational Linguistics.  
+Naman Goyal, Cynthia Gao, Vishrav Chaudhary, Peng-Jen Chen, Guillaume Wenzek, Da Ju, Sanjana Krishnan, Marc'Aurelio Ranzato, Francisco Guzmán, and Angela Fan. 2021. The flores-101 evaluation benchmark for low-resource and multilingual machine translation.  
+Thanh-Le Ha, Jan Niehues, and Alex Waibel. 2016. Toward multilingual neural machine translation with universal encoder and decoder. In Proceedings of the 13th International Conference on Spoken Language Translation, Seattle, Washington D.C. International Workshop on Spoken Language Translation.  
+Taku Kudo. 2018. Subword regularization: Improving neural network translation models with multiple subword candidates. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 66-75, Melbourne, Australia. Association for Computational Linguistics.  
+Taku Kudo and John Richardson. 2018. SentencePiece: A simple and language independent subword tokenizer and tokenizer for neural text processing. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 66-71, Brussels, Belgium. Association for Computational Linguistics.  
+Antonis Maronikolakis, Philipp Dufter, and Hinrich Schütze. 2021. Wine is not v i n. on the compatibility of tokenizations across languages. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 2382-2399, Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Francois Meyer and Jan Buys. 2023. Subword segmental machine translation: Unifying segmentation and target sentence generation. In *Findings of the Association for Computational Linguistics: ACL* 2023, pages 2795-2809, Toronto, Canada. Association for Computational Linguistics.  
+Vaidehi Patil, Partha Talukdar, and Sunita Sarawagi. 2022. Overlap-based vocabulary generation improves cross-lingual transfer among related languages. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 219–233, Dublin, Ireland. Association for Computational Linguistics.  
+Telmo Pires, Eva Schlinger, and Dan Garrette. 2019. How multilingual is multilingual BERT? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4996-5001, Florence, Italy. Association for Computational Linguistics.  
+Maja Popovic. 2017.  $\mathrm{chrF}++$ : words helping character n-grams. In Proceedings of the Second Conference on Machine Translation, pages 612-618, Copenhagen, Denmark. Association for Computational Linguistics.
+
+Rigardt Pretorius, Ansu Berg, Laurette Pretorius, and Biffie Viljoen. 2009. Setswana tokenisation and computational verb morphology: Facing the challenge of a disjunctive orthography. In Proceedings of the First Workshop on Language Technologies for African Languages, pages 66-73, Athens, Greece. Association for Computational Linguistics.  
+Phillip Rust, Jonas Pfeiffer, Ivan Vulic, Sebastian Ruder, and Iryna Gurevych. 2021. How good is your tokenizer? on the monolingual performance of multilingual language models. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 3118-3135, Online. Association for Computational Linguistics.  
+Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Neural machine translation of rare words with subword units. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1715-1725, Berlin, Germany. Association for Computational Linguistics.  
+Uri Shaham, Maha Elbayad, Vedanuj Goswami, Omer Levy, and Shruti Bhosale. 2023. Causes and cures for interference in multilingual translation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15849-15863, Toronto, Canada. Association for Computational Linguistics.  
+David Stap, Vlad Niculae, and Christof Monz. 2023. Viewing knowledge transfer in multilingual machine translation through a representational lens. In *Findings of the Association for Computational Linguistics: EMNLP* 2023, pages 14973-14987, Singapore. Association for Computational Linguistics.  
+Elsabé Taljard and Sonja E. Bosch. 2006. A comparison of approaches to word class tagging: Disjunctively vs. conjunctively written bantu languages. Nordic Journal of African Studies, 15:428-442.  
+NLLB Team, Marta R. Costa-jussà, James Cross, Onur Celebi, Maha Elbayad, Kenneth Heafield, Kevin Hef fernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, Anna Sun, Skyler Wang, Guillaume Wenzek, Al Youngblood, Bapi Akula, Loic Barrault, Gabriel Mejia Gonzalez, Prangthip Hansanti, John Hoffman, Semarley Jarrett, Kaushik Ram Sadagopan, Dirk Rowe, Shannon Spruit, Chau Tran, Pierre Andrews, Necip Fazil Ayan, Shruti Bhosale, Sergey Edunov, Angela Fan, Cynthia Gao, Vedanuj Goswami, Francisco Guzmán, Philipp Koehn, Alexandre Mourachko, Christophe Ropers, Safiyyah Saleem, Holger Schwenk, and Jeff Wang. 2022. No language left behind: Scaling humancentered machine translation.  
+Changhan Wang, Kyunghyun Cho, and Jiatao Gu. 2020a. Neural machine translation with byte-level subwords. In AAAI 2020 - 34th AAAI Conference on
+
+Artificial Intelligence, AAAI 2020 - 34th AAAI Conference on Artificial Intelligence, pages 9154-9160. AAAI press. Publisher Copyright: Copyright © 2020 Association for the Advancement of Artificial Intelligence. All rights reserved.; 34th AAAI Conference on Artificial Intelligence, AAAI 2020; Conference date: 07-02-2020 Through 12-02-2020.
+
+Xinyi Wang, Sebastian Ruder, and Graham Neubig. 2021. Multi-view subword regularization. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 473-482, Online. Association for Computational Linguistics.
+
+Zihan Wang, Karthikeyan K, Stephen Mayhew, and Dan Roth. 2020b. Extending multilingual BERT to low-resource languages. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 2649-2656, Online. Association for Computational Linguistics.
+
+Shijie Wu and Mark Dredze. 2019. Beto, bentz, becas: The surprising cross-lingual effectiveness of BERT. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 833–844, Hong Kong, China. Association for Computational Linguistics.
+
+Judit Ács. 2019. Exploring BERT's vocabulary.
+
+# A Model Configurations
+
+We use the model size and training hyperparameters of the fairseq transformer-base architecture. We train our models for 45 epochs and use a language sampling temperature of  $T = 1.5$  to balance exposure to low-resource and high-resource languages. In our cross-lingual finetuning experiments we finetune models on en→ss for 20 epochs and on en→xh/af/ts for 10 epochs.
+
+When applying subword methods we specify a shared vocabulary size of 8k. This is slightly smaller than the optimal vocabulary size of 10k used by Meyer and Buys (2023) in experiments on these same languages, since we use smaller subsets of the WMT22 datasets.

2024/A Systematic Analysis of Subwords and Cross-Lingual Transfer in Multilingual Translation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:192252ce59b9a596aa236660a656ecce40baae8dff8f45c27a9f96ef02b28179
+size 174877

2024/A Transformer with Stack Attention/a415eb91-fbbd-49b8-a99a-1ef3cad21431_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5c7db821743a61f7cc156424a528fa5086e2e7e9bb866430a9d8d62878bce5aa
+size 661246

2024/A Transformer with Stack Attention/full.md

@@ -0,0 +1,687 @@
+# A Transformer with Stack Attention
+
+Jiaoda Li $^{1}$  Jennifer C. White $^{2}$  Mrinmaya Sachan $^{1}$  Ryan Cotterell $^{1}$
+
+$^{1}$ ETH Zürich  $^{2}$ University of Cambridge
+
+{jiaoda.li, mrinmaya.sachan, ryan.cotterell}@inf.ethz.ch jw2088@cam.ac.uk
+
+# Abstract
+
+Natural languages are believed to be (mildly) context-sensitive. Despite underpinning remarkably capable large language models, transformers are unable to model many context-free language tasks. In an attempt to address this limitation in the modeling power of transformer-based language models, we propose augmenting them with a differentiable, stack-based attention mechanism. Our stack-based attention mechanism can be incorporated into any transformer-based language model and adds a level of interpretability to the model. We show that the addition of our stack-based attention mechanism enables the transformer to model some, but not all, deterministic context-free languages.
+
+https://github.com/rycolab/ stack-transformer
+
+# 1 Introduction
+
+Language models (LMs) based on the transformer architecture (Vaswani et al., 2017) have shown great empirical success at a wide range of NLP tasks (Devlin et al., 2019; Radford et al., 2019; Liu et al., 2020; Brown et al., 2020). However, recent theoretical (Hahn, 2020; Angluin et al., 2023) and empirical (Ebrahimi et al., 2020; Bhattamishra et al., 2020; Deletang et al., 2023) research suggests that language models based on transformers show difficulty in learning basic algorithmic patterns. A prime example is the Dyck- $n$  language, i.e., the language of balanced parentheses of depth  $\leqslant n$ . When  $n > 1$ , it has been argued that transformers are theoretically (Hahn, 2020) and empirically (Ebrahimi et al., 2020) unable to learn a Dyck- $n$  language. Additionally, Deletang et al. (2023) report that transformer-based LMs fail to learn four deterministic context-free (DCF) tasks. The authors of this work contend that the resolution of this insufficiency is paramount if human-level language understanding is to be achieved by computers. Indeed, Chomsky (1956) famously argues that human language has many context-free traits; see also Chomsky and Schützenberger (1963). Moreover, Shieber (1987)
+
+goes further and argues that snippets of Swiss German are even higher on the Chomsky hierarchy.
+
+The scientific question treated in this paper is whether there exists a minimal modification to the transformer architecture that does allow it to learn a larger swathe of the formal languages most closely associated with human language. Specifically, in this paper, we augment the transformer architecture with a novel stack attention mechanism that enables it to learn certain CF languages empirically. Our stack attention mechanism simulates a stack by maintaining a probability distribution over which of the subsequently observed tokens is at the top element of the stack. In turn, this probability distribution serves as an attention mechanism. Compared to DuSell and Chiang (2024), which also applies stack augmentation to the transformer, our stack attention is more space efficient and allows for easier interpretation through visualizing the attention weights. We incorporate our innovation into the transformer by adding a stack attention sub-layer to each layer, rather than completely replacing the standard attention. Augmenting models in a modular way like this allows for direct integration with pre-trained transformer-based LMs.
+
+We evaluate our stack-augmented transformer through comparison with a standard transformer on four DCF tasks taken from Delétang et al. (2023). We find that the stack-augmented transformer performs substantially better than the standard transformer on two of the four DCF tasks. Nevertheless, in contrast to DuSell and Chiang (2024), who claim their architecture can recognize the entire class of CF languages, we find transformers with our stack attention still struggle on two DCF tasks that involve modular arithmetic.
+
+# 2 Preliminaries
+
+In this section, we provide the necessary technical background for our exposition. We first review the self-attention mechanism and then introduce the transformer architecture.
+
+# 2.1 The Self-Attention Mechanism
+
+The attention mechanism (Bahdanau et al., 2015) is the fundamental building block of the trans
+
+former architecture (Vaswani et al., 2017), which we discuss in the next section. One common form of attention is self-attention (Cheng et al., 2016; Parikh et al., 2016). Our construction of a stack-augmented attention mechanism is a modification of this self-attention mechanism.
+
+The premise of self-attention is as follows. A sentence representation  $\mathbf{H} = [\mathbf{h}_1; \ldots; \mathbf{h}_N] \in \mathbb{R}^{D \times N}$  is a horizontal concatenation of column vectors  $\mathbf{h}_n$  in  $\mathbb{R}^D$ , where each column is a representation of the  $n^{\text{th}}$  word. Our goal is to construct a distribution over the index set  $\{1, \ldots, N\}$ , denoted as  $[N]$ . We do so in three steps, described below.
+
+① The first step is to construct a real-valued, pairwise compatibility score. The most common way to do this is through a (scaled) dot-product, i.e.,
+
+$$
+e _ {i j} \stackrel {\text {d e f}} {=} \frac {\mathbf {h} _ {i} \cdot \mathbf {h} _ {j}}{\sqrt {D}} \tag {1}
+$$
+
+(2) The second step is to take the pair-wise compatibility scores and project them onto the simplex  $\Delta^{N-1}$  through the softmax. This results in the following distribution
+
+$$
+\boldsymbol {\alpha} _ {i} (j) \stackrel {\text {d e f}} {=} \frac {\exp \left(e _ {i j}\right)}{\sum_ {n = 1} ^ {N} \exp \left(e _ {i n}\right)} \tag {2}
+$$
+
+which is termed the self-attention distribution. Note there are  $N$  self-attention distributions  $\alpha_{i}$
+
+The third, and final, step is to construct a weighted average of the representations  $\mathbf{H} = [\mathbf{h}_1; \ldots; \mathbf{h}_N] \in \mathbb{R}^{D \times N}$  using the self-attention distribution as follows
+
+$$
+\mathbf {A} (\mathbf {H}) _ {:, i} \stackrel {\text {d e f}} {=} \sum_ {n = 1} ^ {N} \boldsymbol {\alpha} _ {i} (n) \mathbf {h} _ {n} \tag {3}
+$$
+
+where  $\mathbf{A}(\mathbf{H})_{:,i}$  denotes the  $i^{\mathrm{th}}$  column of  $\mathbf{A}(\mathbf{H})$ . The function  $\mathbf{A}:\mathbb{R}^{D\times N}\to \mathbb{R}^{D\times N}$  (for any  $N$ ), as defined above, is called an attention head.
+
+Importantly, we see that  $\mathbf{A}$  is a differentiable function. Differentiability allows us to learn the parameters of an attention head with gradient-based methods. And, more importantly, it has a specific desirable property—namely, it is invariant with respect to permutations of the columns of  $\mathbf{H}$ . Computationally, this implies that  $\mathbf{A}(\mathbf{H})_{:,i}$  and  $\mathbf{A}(\mathbf{H})_{:,j}$  can be computed in parallel for  $i \neq j$ . It is specifically this form of parallelism that grants the transformer architecture its ability to scale.
+
+One drawback of the permutation invariance, however, is that  $\mathbf{A}$  is not a linguistically plausible mechanism as human language is decidedly not permutation invariant. This problem is addressed through the incorporation of attention masks and positional encodings (Vaswani et al., 2017, § 3.5) in the transformer architecture, as we discuss in §2.2.
+
+Masked Self-Attention. An attention mask  $\mathbf{G} \in \mathbb{B}^{N \times N}$ , where  $\mathbb{B} = \{0,1\}$ , can be applied to the self-attention using the following generalization
+
+$$
+\boldsymbol {\alpha} _ {i} (j) = \frac {\exp \left(e _ {i j}\right) \mathbf {G} _ {i , j}}{\sum_ {n = 1} ^ {N} \exp \left(e _ {i n}\right) \mathbf {G} _ {i , n}} \tag {4}
+$$
+
+Attention masks allow for hard constraints on which indices can be attended to by the attention head. A commonly used masking scheme is future masking where each position is only allowed to attend to positions up to and including itself, i.e., we define the following mask
+
+$$
+\mathbf {G} _ {i, n} = \left\{ \begin{array}{l l} 1, & n <   i \\ 0, & n \geqslant i \end{array} \right. \tag {5}
+$$
+
+Future masking allows transformers to be used in autoregressive language models by preventing the model from peeking at words that have yet to be generated, which we detail in §2.3.
+
+Queries, Keys, and Values. In the version of the attention mechanism introduced by Vaswani et al. (2017), the attention mechanism is augmented with additional linear projections. Specifically, the vectors  $\mathbf{h}_n$  are linearly projected to construct three new vectors, defined below
+
+$$
+\mathbf {q} _ {n} \stackrel {\text {d e f}} {=} \mathbf {W} _ {\mathrm {Q}} \mathbf {h} _ {n} \quad \tag {6a}
+$$
+
+$$
+\mathbf {k} _ {n} \stackrel {\text {d e f}} {=} \mathbf {W} _ {\mathrm {K}} \mathbf {h} _ {n} \quad (\text {k e y}) \tag {6b}
+$$
+
+$$
+\mathbf {v} _ {n} \stackrel {\text {d e f}} {=} \mathbf {W} _ {\mathrm {V}} \mathbf {h} _ {n} \quad (\text {v a l u e}) \tag {6c}
+$$
+
+where  $\mathbf{W}_{\mathrm{V}},\mathbf{W}_{\mathrm{Q}},\mathbf{W}_{\mathrm{K}}\in \mathbb{R}^{D^{\prime}\times D}$  are parameter matrices. Compatibility scores are then computed between the corresponding query-key pair:
+
+$$
+e _ {i j} = \frac {\mathbf {q} _ {i} \cdot \mathbf {k} _ {j}}{\sqrt {D ^ {\prime}}} \tag {7}
+$$
+
+Using those compatibility scores, a self-attention distribution is constructed using the softmax. Finally, as before, a weighted sum of the values is computed using the self-attention distribution:
+
+$$
+\mathbf {A} (\mathbf {H}) _ {:, i} = \sum_ {n = 1} ^ {N} \boldsymbol {\alpha} _ {i} (n) \mathbf {v} _ {n} \tag {8}
+$$
+
+Multi-head Self-Attention. We additionally define the multi-head self-attention mechanism. In multi-head attention, we combine  $M$  attention heads  $\mathbf{A}^{(1)},\ldots ,\mathbf{A}^{(M)}$  as follows
+
+$$
+\mathbf {M} (\mathbf {H}) _ {:, i} \stackrel {\text {d e f}} {=} \sum_ {m = 1} ^ {M} \mathbf {W} _ {\mathrm {O}} ^ {(m)} \mathbf {A} ^ {(m)} (\mathbf {H}) _ {:, i} \tag {9}
+$$
+
+where  $\mathbf{W}_{\mathrm{O}}^{(m)}\in \mathbb{R}^{D\times D^{\prime}}$  is the output projection matrix for head  $\mathbf{A}^{(m)}$ . Usually, we set  $D^{\prime} = D / M$ .
+
+# 2.2 The Transformer Architecture
+
+We now describe the transformer architecture. A transformer over a vocabulary  $\Sigma$  constitutes a function of type $^{1}$ $\Sigma^{N} \to \mathbb{R}^{D \times N}$  where a string  $\boldsymbol{w} = w_{1} \cdots w_{N} \in \Sigma^{N}$  of length  $N$  is encoded into a  $\mathbb{R}^{D \times N}$  representation where  $D$  is the model size. The transformer is defined compositionally over a sequence of layers. First, we define
+
+$$
+\mathbf {H} ^ {(0)} \stackrel {\text {d e f}} {=} \text {E m b e d d i n g} + \mathrm {P E} \tag {10}
+$$
+
+where Embedding of type  $\Sigma^N\to \mathbb{R}^{D\times N}$  is the embedding layer and PE of type  $\Sigma^N\rightarrow \mathbb{R}^{D\times N}$  is the positional encoding that injects information about the relative or absolute position of tokens, which extinguishes the permutation invariance of the transformer. Each transformer layer consists of two sub-layers: a multi-head self-attention  $\mathbf{M}$  of type  $\mathbb{R}^{D\times N}\to \mathbb{R}^{D\times N}$  and a fully connected feedforward network FFN of type  $\mathbb{R}^{D\times N}\to \mathbb{R}^{D\times N}$ . A residual connection is employed around each sublayer, followed by a layer normalization (Ba et al., 2016) LN of type  $\mathbb{R}^{D\times N}\to \mathbb{R}^{D\times N}$ : for  $0 < \ell \leqslant L$ , we have the following recursive definition
+
+$$
+\mathbf {H} _ {\mathbf {M}} ^ {(\ell)} \stackrel {\text {d e f}} {=} \operatorname {L N} \left(\mathbf {M} \left(\mathbf {H} ^ {(\ell - 1)}\right) + \mathbf {H} ^ {(\ell - 1)}\right) \tag {11a}
+$$
+
+$$
+\mathbf {H} _ {\mathrm {F F N}} ^ {(\ell)} \stackrel {\text {d e f}} {=} \operatorname {L N} \left(\operatorname {F F N} \left(\mathbf {H} _ {\mathbf {M}} ^ {(\ell)}\right) + \mathbf {H} _ {\mathbf {M}} ^ {(\ell)}\right) \tag {11b}
+$$
+
+$$
+\mathbf {H} ^ {(\ell)} \stackrel {\text {d e f}} {=} \mathbf {H} _ {\mathrm {F F N}} ^ {(\ell)} \tag {11c}
+$$
+
+where  $\mathbf{H}_{\mathbf{M}}^{(\ell)}$ ,  $\mathbf{H}_{\mathrm{FFN}}^{(\ell)}$  and  $\mathbf{H}^{(\ell)}$  are functions of type  $\Sigma^N \to \mathbb{R}^{D \times N}$  for any  $N$ . They have  $\boldsymbol{w}$  as input, which we omit for brevity when the context is clear.
+
+Future-masked Transformer. If the future mask in Eq. (5) is used in every  $\mathbf{H}_{\mathbf{M}}^{(\ell)}$ , we call such a transformer future-masked transformer, denoted as  $\mathbf{F}^{(L)}$ . As we will see, future-masked transformers are necessary to construct autoregressive language models, which cannot peek at the future.
+
+$^{1}$ Type-theoretically,  $N$  is a parameter of the type. Thus, our type signature is a dependent type (Univalent Foundations Program, 2013)
+
+# 2.3 Probability Models
+
+Next, we describe two natural ways of constructing a probability distribution from a transformer.
+
+Masked Language Modeling. First, we consider the case of masked language modeling (MLM). Masked language models perform the cloze task, i.e., they fill in a missing word given a left and right context. More formally, consider a string  $\boldsymbol{w} \in \Sigma^{*}$  of length  $T$ . Let  $w_{t}$  denote the  $t^{\text{th}}$  symbol in  $\boldsymbol{w}$ , let  $\boldsymbol{w}_{<t} = w_{1} \cdots w_{t-1}$ , and let  $\boldsymbol{w}_{>t} = w_{t+1} \cdots w_{T}$ . We construct  $\widetilde{\boldsymbol{w}} \stackrel{\text{def}}{=} \boldsymbol{w}_{<t}[\text{MASK}]\boldsymbol{w}_{>t}$  by replacing  $w_{t}$  in  $\boldsymbol{w}$  with a mask token [MASK]. The alphabet is expanded to include [MASK]. We denote  $\widetilde{\Sigma} \stackrel{\text{def}}{=} \Sigma \cup \{[\text{MASK}]\}$ . The transformer  $\mathbf{H}^{(L)}$  is now of type  $\widetilde{\Sigma}^{N} \to \mathbb{R}^{D \times N}$ . A masked language gives the following probability distribution for position  $t$
+
+$$
+\begin{array}{r l} p \left(\widetilde {w} _ {t} \mid \boldsymbol {w} _ {<   t}, \boldsymbol {w} _ {> t}\right) & \\ = \operatorname {s o f t m a x} \left(\mathbf {W} _ {\mathrm {P}} \mathbf {H} ^ {(L)} \left(\widetilde {\boldsymbol {w}}\right) _ {:, t} + \mathbf {b} _ {\mathrm {P}}\right) _ {\widetilde {w} _ {t}} \end{array} \tag {12}
+$$
+
+where  $\tilde{w}_t\in \tilde{\Sigma}$ $\mathbf{W}_{\mathrm{P}}\in \mathbb{R}^{|\tilde{\Sigma} | \times D}$  and  $\mathbf{b}_{\mathrm{P}}\in \mathbb{R}^{|\tilde{\Sigma} |}$ . In practice, multiple tokens may be masked and predicted simultaneously.
+
+Autoregressive Language Modeling. In contrast to masked language modeling, the goal of autoregressive language modeling (ALM) is to construct a probability distribution over  $\Sigma^{*}$ . To do so, the following factorization is employed
+
+$$
+p (\boldsymbol {w}) = p ([ \mathrm {E O S} ] \mid \boldsymbol {w}) \prod_ {t = 1} ^ {T} p \left(w _ {t} \mid \boldsymbol {w} _ {<   t}\right) \tag {13}
+$$
+
+Every local conditional distribution  $p(w_{t} \mid \boldsymbol{w}_{<t})$  is defined over the set  $\overline{\Sigma} \stackrel{\mathrm{def}}{=} \Sigma \cup \{[\mathrm{EOS}]\}$  and  $\boldsymbol{w}_{<1} \stackrel{\mathrm{def}}{=} [\mathrm{BOS}]$ , where [BOS], [EOS]  $\notin \Sigma$ . In a transformer-based autoregressive language model, each local conditional is parameterized as
+
+$$
+\begin{array}{l} p \left(\bar {w} _ {t} \mid \boldsymbol {w} _ {<   t}\right) \tag {14} \\ = \operatorname {s o f t m a x} \left(\mathbf {W} _ {\mathrm {P}} \mathbf {F} ^ {(L)} (\boldsymbol {w}) _ {:, t - 1} + \mathbf {b} _ {\mathrm {P}}\right) _ {\overline {{\boldsymbol {w}}} _ {t}} \\ \end{array}
+$$
+
+where  $\overline{w}_t\in \overline{\Sigma}$ $\mathbf{F}^{(L)}$  is a future-masked transformer,  $\mathbf{W}_{\mathrm{P}}\in \mathbb{R}^{|\overline{\Sigma} | \times D}$  and  $\mathbf{b}_{\mathrm{P}}\in \mathbb{R}^{|\overline{\Sigma} |}$ .
+
+# 3 A Transformer with Stack Attention
+
+Recently Delétang et al. (2023) showed that transformers are not able to learn several non-regular
+
+<table><tr><td>Action</td><td colspan="3">Stack</td><td colspan="3">Attention</td><td>Stack Attention α</td></tr><tr><td></td><td></td><td></td><td></td><td>[BOS]a↑</td><td>b</td><td>c</td><td>α0=[1,0,0,0]T</td></tr><tr><td>PUSH a</td><td>a</td><td></td><td></td><td>[BOS]a↑</td><td>b</td><td>c</td><td>α1=[0,1,0,0]T</td></tr><tr><td>PUSH b</td><td>b</td><td>a</td><td></td><td>[BOS]a</td><td>b</td><td>c</td><td>α2=[0,0,1,0]T</td></tr><tr><td>PUSH c</td><td>c</td><td>b</td><td>a</td><td></td><td>b</td><td>c</td><td>α3=[0,0,0,1]T</td></tr><tr><td>POP</td><td>b</td><td>a</td><td></td><td>[BOS]a</td><td>b</td><td>c</td><td>α4=∑j=13α3(j)αj-1=α2=[0,0,1,0]T</td></tr><tr><td>NO-OP</td><td>b</td><td>a</td><td></td><td>[BOS]a</td><td>b</td><td>c</td><td>α5=α4=[0,0,1,0]T</td></tr><tr><td>POP</td><td>a</td><td></td><td></td><td>[BOS]a</td><td>b</td><td>c</td><td>α6=∑j=15α5(j)αj-1=α1=[0,1,0,0]T</td></tr></table>
+
+Figure 1: An example illustrating how attentions can emulate stacks. The first column lists the operation performed at each timestep. The second column presents the stack contents after performing the operation. The third column shows a hard attention over the input tokens. The pointer of the attention indicates the current stack top. The last column is the proposed stack attention.
+
+DCF languages. Inspired by the fact that pushdown automata (Oettinger, 1961; Schützenberger, 1963), automata that employ a single stack, can model CF languages (Evey, 1963), we introduce a novel stack attention mechanism that emulates the functionality of a stack and integrate it into the transformer architecture, aiming to enable it to learn some CF languages.
+
+# 3.1 Stacks over the Index Set
+
+We first give a formal definition of a stack. In our paper, we define a stack as a data structure over the index set  $[N]$ . The state of a stack is a string  $\gamma \in [N]^*$  of indices. There are three operations that we can perform that alter the state of the stack. We describe each operation below in terms of  $\gamma$ .
+
+- The operation  $\mathsf{PUSH}\colon [N]^*\times [N]\to [N]^*$  adds an element to the top of the stack and is formally defined as follows:
+
+$$
+\operatorname {P U S H} (\gamma , \gamma) = \gamma \gamma \tag {15}
+$$
+
+- The operation  $\mathsf{NO - OP}\colon [N]^*\to [N]^*$  leaves the stack unchanged and is defined as follows:
+
+$$
+\mathrm {N O} - \mathrm {O P} (\gamma) = \gamma \tag {16}
+$$
+
+- The operation POP:  $[N]^* \to [N]^*$  removes the top-most element of the stack and is formally defined as follows:
+
+$$
+\operatorname {P O P} (\varepsilon) = \varepsilon \tag {17a}
+$$
+
+$$
+\operatorname {P O P} \left(\gamma_ {1} \dots \gamma_ {T}\right) = \gamma_ {1} \dots \gamma_ {T - 1} \tag {17b}
+$$
+
+We will use this definition in Theorem 3.1 to argue that our stack attention mechanism can be formally viewed as a type of stack. Additionally, we will assume an operator PEEK:  $[N]^* \to ([N] \cup \{0\})$  that does not alter the state of the stack, but rather returns the top element (or 0 if the stack is empty). We define it below
+
+$$
+\mathrm {P E E K} (\varepsilon) = 0 \tag {18a}
+$$
+
+$$
+\mathrm {P E E K} \left(\gamma_ {1} \dots \gamma_ {T}\right) = \gamma_ {T} \tag {18b}
+$$
+
+# 3.2 Stack Attention
+
+We now formally define our stack attention mechanism. We introduce a beginning-of-sequence symbol [BOS] at the zeroth position, designated to represent an empty stack. Each position  $i \in \{0\} \cup [N]$  is assigned a distinct stack  $\alpha_{i} \in \mathbb{R}^{N + 1}$ . We write  $\alpha_{i}(j)$  to denote the  $(j + 1)^{\mathrm{th}}$  value in  $\alpha_{i}$ , for  $0 \leqslant j \leqslant N$ . The stacks are defined inductively.
+
+The initial stack,  $\alpha_0$ , is constructed to attend to [BOS] as follows
+
+$$
+\boldsymbol {\alpha} _ {0} \stackrel {\text {d e f}} {=} [ 1, 0, 0, \dots ] ^ {\top} \in \mathbb {R} ^ {N + 1} \tag {19}
+$$
+
+Subsequent stacks are computed inductively based on the stack contents and the operations (PUSH, NO-OP, POP) taken at previous timesteps. The three stack operations are defined for  $i \geqslant 1$  as follows:
+
+- PUSH pushes the hidden state of the current position, so we just set the attention weight at the current position to be 1 and the rest to be 0, i.e.,
+
+$$
+\boldsymbol {\alpha} _ {i} ^ {(\mathrm {P U S H})} (j) \stackrel {\text {d e f}} {=} \left\{ \begin{array}{l l} 1 & j = i \\ 0 & \text {o t h e r w i s e} \end{array} \right. \tag {20}
+$$
+
+- NO-OP leaves the previous stack unchanged, so the stack from the last timestep is passed forward with no modification, i.e., we have
+
+$$
+\boldsymbol {\alpha} _ {i} ^ {(N 0 - O P)} \stackrel {\text {d e f}} {=} \boldsymbol {\alpha} _ {i - 1} \tag {21}
+$$
+
+- POP removes the top element, and backtracks to the second element in the stack, i.e., we have
+
+$$
+\boldsymbol {\alpha} _ {i} ^ {(P O P)} \stackrel {\text {d e f}} {=} \left[ \sum_ {j = 1} ^ {i - 1} \boldsymbol {\alpha} _ {i - 1} (j) \boldsymbol {\alpha} _ {j - 1} \right] + \boldsymbol {\alpha} _ {i - 1} (0) \boldsymbol {\alpha} _ {0} \tag {22}
+$$
+
+The first term on the right-hand side retrieves the second element and is zeroed out when  $i = 1$ . The second term accounts for the case of an empty stack—POP cannot be performed on an empty stack and in this case it is equivalent to a NO-OP.
+
+The stack attention  $\alpha_{i}$  at position  $i$  is computed as a superposition of the three operations:
+
+$$
+\begin{array}{l} \boldsymbol {\alpha} _ {i} \stackrel {\text {d e f}} {=} \mathbf {a} _ {i} (\text {P U S H}) \cdot \boldsymbol {\alpha} _ {i} ^ {(\text {P U S H})} + \mathbf {a} _ {i} (\text {P O P}) \cdot \boldsymbol {\alpha} _ {i} ^ {(\text {P O P})} \tag {23} \\ + \mathbf {a} _ {i} (\mathrm {N O} - \mathrm {O P}) \cdot \boldsymbol {\alpha} _ {i} ^ {(\mathrm {N O} - \mathrm {O P})} \\ \end{array}
+$$
+
+where  $\mathbf{a}_i\in \Delta^2$  is a probability distribution over possible operations  $\mathcal{A} = \{\mathrm{PUSH},\mathrm{POP},\mathrm{NO - OP}\}$ . This distribution is determined by:
+
+$$
+\mathbf {a} _ {i} \stackrel {\text {d e f}} {=} \operatorname {s o f t m a x} \left(\mathbf {W} _ {\mathrm {A}} \mathbf {h} _ {i} + \mathbf {b} _ {\mathrm {A}}\right) \tag {24}
+$$
+
+where  $\mathbf{W}_{\mathrm{A}}\in \mathbb{R}^{3\times D}$  and  $\mathbf{b}_{\mathrm{A}}\in \mathbb{R}^{3}$  are learned parameters.
+
+After obtaining the stack attention weights, we can compute the top element as a weighted sum just like standard self-attention:
+
+$$
+\mathbf {S} (\mathbf {H}) _ {:, i} \stackrel {\mathrm {d e f}} {=} \sum_ {n = 0} ^ {N} \boldsymbol {\alpha} _ {i} (n) \mathbf {h} _ {n} \tag {25}
+$$
+
+# 3.3 A Stack Transformer
+
+The stack is incorporated into the transformer by inserting a third sub-layer in each transformer layer after the standard attention and feedforward layers defined in Eq. (11a) and Eq. (11b):
+
+$$
+\mathbf {H} _ {\mathbf {S}} ^ {(\ell)} \stackrel {\text {d e f}} {=} \mathbf {S} \left(\mathbf {H} _ {\mathrm {F F N}} ^ {(\ell)}\right) + \mathbf {H} _ {\mathrm {F F N}} ^ {(\ell)} \tag {26a}
+$$
+
+$$
+\mathbf {H} ^ {(\ell)} = \mathbf {H} _ {\mathbf {S}} ^ {(\ell)} \tag {26b}
+$$
+
+Similar to other sub-layers, we also employ a residual connection by summing the output of the stack attention mechanism S with its input, allowing the model to bypass the stack if needed. Layer normalization can also be used, but we omit due to initial results in preliminary experiments. Because the rest of the model is left unchanged, it can be directly integrated into pre-trained language models to augment their ability to process hierarchical syntactic structures.
+
+# 3.4 Computational Overhead
+
+Time. The computation is bottlenecked by the POP operation, which sums over previous the previous positions and thereby has a time complexity of  $\mathcal{O}(N)$ . The total time complexity is  $\mathcal{O}\left(N^{2}\right)$ . In contrast to standard attention, stack attention has to be computed sequentially, which breaks the parallelizability of the transformer and makes it substantially slower in practice. However, a and the output  $\mathbf{S}(\mathbf{H})$  can still be computed in parallel. Thus,  $\alpha$  is a function of the stack operations but not of the hidden states. It follows that if structural supervision of the stack operations is provided, e.g., as in Sartran et al. (2022) and Murty et al. (2023),  $\alpha_{i}$  for all  $i\in [N]$  can be pre-computed, and the entire model can be parallelized.
+
+Space. The stack attention stores  $N + 1$  attentions of size  $N$ , so the space complexity is  $\mathcal{O}\left((N + 1)N\right) = \mathcal{O}\left(N^2\right)$ . This is an improvement over the  $\mathcal{O}\left(DN^2\right)$  space complexity of DuSell and Chiang's (2024) method.
+
+# 3.5 The Duality of Stack Attention
+
+Stack attention is both a stack over the index set, as defined in §3.1, and an attention mechanism. In the following theorem, we make precise the manner in which our stack attention is a stack.
+
+Notation. We use the symbol  $\upsilon_{i}$  to refer to an operation from the set  $\{\mathsf{PUSH}_i(\cdot),\mathsf{NO - OP}(\cdot),\mathsf{POP}(\cdot)\}$  at every time step  $i$ . Note that PUSH, as defined in
+
+<table><tr><td rowspan="2">Task</td><td colspan="2">RNN</td><td colspan="2">Transformer MLM</td><td colspan="2">Transformer ALM</td></tr><tr><td>Vanilla</td><td>Stack</td><td>Vanilla</td><td>Stack</td><td>Vanilla</td><td>Stack</td></tr><tr><td>RS</td><td>81.0 ± 0.8</td><td>100.0 ± 0.0</td><td>54.8 ± 0.0</td><td>100.0 ± 0.0</td><td>55.4 ± 0.8</td><td>100.0 ± 0.0</td></tr><tr><td>SM</td><td>73.2 ± 1.0</td><td>100.0 ± 0.0</td><td>50.4 ± 0.1</td><td>93.1 ± 4.4</td><td>50.4 ± 0.1</td><td>92.8 ± 2.6</td></tr><tr><td>MA</td><td>75.8 ± 4.3</td><td>91.0 ± 6.3</td><td>30.1 ± 0.0</td><td>34.3 ± 1.4</td><td>30.2 ± 0.1</td><td>29.5 ± 0.6</td></tr><tr><td>SE</td><td>56.7 ± 10.3</td><td>89.9 ± 7.2</td><td>20.0 ± 0.0</td><td>29.8 ± 8.0</td><td>20.2 ± 0.1</td><td>20.3 ± 0.2</td></tr></table>
+
+Table 1: Accuracies (%) of the transformer and RNN without and with stacks on DCF tasks.
+
+$\S 3.1$ , is a function of two arguments. However, we define  $\mathsf{PUSH}_i(\gamma) \stackrel{\mathrm{def}}{=} \mathsf{PUSH}(\gamma, i)$ , i.e., we push  $i$ , the index, to the stack. We introduce a function  $\mathbb{I}$  of type  $\{0\} \cup [N] \to \mathbb{B}^{N+1}$  that converts an index into its one-hot encoding, a column vector with zeros everywhere except the given index, where the value is set to one.
+
+Theorem 3.1. Let  $v_{1}, \ldots, v_{N}$  be a series of stack operations where  $v_{i} \in \{\mathsf{PUSH}_{i}(\cdot), \mathsf{NO - OP}(\cdot), \mathsf{POP}(\cdot)\}$  for all  $i \in [N]$ . Furthermore, suppose  $\mathbf{a}_{i}(v_{i}) = 1$  for all  $i \in [N]$ . Then,  $[[\mathsf{PEEK}(v_{i}(\cdots v_{1}(\varepsilon)))]] = \alpha_{i}$  for all  $i \in \{0\} \cup [N]$ .
+
+Proof. Appendix A
+
+Our stack-based attention is also an attention mechanism in the sense that it maintains a distribution over  $\{0\} \cup [N]$ . We make this notion precise as well in the following theorem.
+
+Theorem 3.2. Consider a sequence of stack attention mechanisms  $\alpha_0, \ldots, \alpha_N$ . Then,  $\sum_{n=0}^{N} \alpha_i(n) = 1$  for all  $i \in \{0\} \cup [N]$ .
+
+Proof. Appendix A
+
+# 3.6 Expressivity
+
+We leave the exact characterization of the expressivity of our stack-augmented transformer for future work. However, we conjecture that it cannot model all the context-free languages without positional encodings. Such a result would mirror that of Angluin et al. (2023).
+
+To contextualize this conjecture, we first review the star-free languages. The star-free languages are regular languages definable by a regular expression without Kleene star but with complement (McNaughton and Papert, 1971). They can also be characterized by finite-state automata with aperiodic transformation monoids (Schützenberger, 1965), also termed counter-free automata or permutation-free automata (McNaughton and Papert, 1971). It has been shown that a counter-free automaton can
+
+only perform counting up to a threshold, but not modulo counting (McNaughton and Papert, 1971).
+
+Recently, Angluin et al. (2023) showed that the class of languages recognizable by transformer encoders with hard attention, strict future masking, and no positional encodings, are exactly the star-free languages. Building on this result, we conjecture that there exist non-star-free languages that are beyond the capability of a transformer encoder with (hard) stack attention and no positional encodings. This conjecture is supported by our experiments in §4.3 where we show that stack-augmented transformers also fail to learn two tasks involving modulo counting. We hope to construct a proof of an expressivity result in future work.
+
+# 4 Deterministic CF Tasks
+
+We now discuss several tasks that are encodable by deterministic context-free grammars.
+
+# 4.1 Tasks
+
+All four tasks we consider are derived from Dele-tang et al. (2023) and are language transduction tasks. Every word from the input language  $\pmb{x} \in \Sigma_{\mathrm{I}}^{*}$  is mapped to a word in the output language  $\pmb{y} \in \Sigma_{\mathrm{O}}^{*}$  by means of a function  $f \colon \Sigma_{\mathrm{I}}^{*} \to \Sigma_{\mathrm{O}}^{*}$ . To convert a transduction task to a language acceptance task, a language is constructed over the alphabet  $\Sigma = \Sigma_{\mathrm{I}} \cup \Sigma_{\mathrm{O}}$  as follows
+
+$$
+\left\{\boldsymbol {x} f (\boldsymbol {x}) \mid \boldsymbol {x} \in \Sigma_ {\mathrm {I}} ^ {*} \right\} \subseteq \Sigma^ {*} \tag {27}
+$$
+
+To experiment with this setup, in the case of an MLM, the input  $\widetilde{\boldsymbol{w}}$  is  $\pmb{x}$  appended with  $|\pmb{y}|$  mask tokens [MASK]. We then use the transformer to predict all the masked tokens at once and evaluate the predicted string  $\pmb{y}'$  against  $\pmb{y} = f(\pmb{x})$ . Likewise, in the case of an ALM, given a prefix  $\pmb{x}$ , we sample  $y_{t}' \sim p(\cdot \mid x y_{<t}')$  autoregressively, where  $y_{t}'$  denotes the  $t^{\text{th}}$  symbol of  $\pmb{y}'$  and  $y_{<t}' = y_{1}' \cdots y_{t-1}'$ . As in the case of MLM, we evaluate the predicted
+
+<table><tr><td>Task</td><td>Transformer</td><td>none</td><td>sincos</td><td>relative</td><td>rotary</td><td>ALiBi</td></tr><tr><td rowspan="2">RS</td><td>Vanilla</td><td>54.8 ± 0.0</td><td>50.7 ± 0.2</td><td>67.6 ± 2.2</td><td>55.4 ± 1.2</td><td>79.4 ± 3.5</td></tr><tr><td>Stack</td><td>100.0 ± 0.0</td><td>99.1 ± 1.7</td><td>100.0 ± 0.0</td><td>86.3 ± 15.0</td><td>100.0 ± 0.0</td></tr><tr><td rowspan="2">SM</td><td>Vanilla</td><td>50.4 ± 0.1</td><td>49.5 ± 0.6</td><td>67.5 ± 1.0</td><td>52.1 ± 1.8</td><td>70.9 ± 1.2</td></tr><tr><td>Stack</td><td>93.1 ± 4.4</td><td>74.7 ± 8.8</td><td>98.5 ± 1.1</td><td>73.1 ± 4.5</td><td>92.9 ± 2.7</td></tr><tr><td rowspan="2">MA</td><td>Vanilla</td><td>30.1 ± 0.0</td><td>30.1 ± 0.0</td><td>30.1 ± 0.0</td><td>30.1 ± 0.0</td><td>30.1 ± 0.0</td></tr><tr><td>Stack</td><td>34.3 ± 1.4</td><td>33.8 ± 0.8</td><td>35.0 ± 1.1</td><td>34.5 ± 1.3</td><td>34.7 ± 1.1</td></tr><tr><td rowspan="2">SE</td><td>Vanilla</td><td>20.0 ± 0.0</td><td>20.0 ± 0.0</td><td>20.0 ± 0.0</td><td>20.0 ± 0.0</td><td>20.0 ± 0.0</td></tr><tr><td>Stack</td><td>29.8 ± 8.0</td><td>23.9 ± 3.0</td><td>25.2 ± 1.8</td><td>30.0 ± 3.8</td><td>27.9 ± 5.8</td></tr></table>
+
+(a) MLM  
+
+<table><tr><td>Task</td><td>Transformer</td><td>none</td><td>sincos</td><td>relative</td><td>rotary</td><td>ALiBi</td></tr><tr><td rowspan="2">RS</td><td>Vanilla</td><td>55.4 ± 0.8</td><td>55.2 ± 0.7</td><td>62.0 ± 6.1</td><td>72.9 ± 3.5</td><td>57.1 ± 0.6</td></tr><tr><td>Stack</td><td>100.0 ± 0.0</td><td>96.8 ± 4.5</td><td>100.0 ± 0.0</td><td>100.0 ± 0.0</td><td>100.0 ± 0.0</td></tr><tr><td rowspan="2">SM</td><td>Vanilla</td><td>64.9 ± 2.0</td><td>60.8 ± 3.1</td><td>70.5 ± 0.9</td><td>71.9 ± 0.9</td><td>70.5 ± 1.6</td></tr><tr><td>Stack</td><td>92.8 ± 2.6</td><td>49.6 ± 4.4</td><td>93.2 ± 2.3</td><td>83.8 ± 1.7</td><td>93.4 ± 1.2</td></tr><tr><td rowspan="2">MA</td><td>Vanilla</td><td>30.2 ± 0.1</td><td>25.7 ± 2.3</td><td>30.3 ± 0.1</td><td>26.0 ± 0.8</td><td>30.3 ± 0.1</td></tr><tr><td>Stack</td><td>30.0 ± 0.1</td><td>28.0 ± 2.8</td><td>30.3 ± 0.3</td><td>25.6 ± 0.3</td><td>30.3 ± 0.1</td></tr><tr><td rowspan="2">SE</td><td>Vanilla</td><td>20.2 ± 0.1</td><td>20.2 ± 0.3</td><td>20.7 ± 0.2</td><td>20.3 ± 0.2</td><td>20.5 ± 0.3</td></tr><tr><td>Stack</td><td>20.3 ± 0.2</td><td>20.2 ± 0.1</td><td>20.7 ± 0.1</td><td>20.2 ± 0.3</td><td>20.3 ± 0.1</td></tr></table>
+
+(b) ALM
+
+Table 2: Performance comparison of a vanilla and stack transformer with different positional encodings.
+
+$\pmb{y}^{\prime}$  against the  $\pmb {y} = f(\pmb {x})$  . We follow the choices of Delétang et al. (2023) for  $\Sigma_{\mathrm{I}}$  and  $\Sigma_0$
+
+Reverse String (RS). The first task is to compute the reverse of an input string, i.e.,  $f_{\mathrm{RS}}(\boldsymbol{x}) = \boldsymbol{x}^R$ . In this task, we take  $\Sigma_{\mathrm{I}} = \Sigma_{\mathrm{O}} = \{\mathsf{a}, \mathsf{b}\}$ . We give an example below.
+
+Example:
+
+$$
+\boldsymbol {x} = \mathrm {a b b}
+$$
+
+$$
+\boldsymbol {y} = \mathrm {b b a}
+$$
+
+Stack Manipulation (SM). In the second task, the input string  $\pmb{x}$  consists of a stack of two symbols  $\{\mathsf{a},\mathsf{b}\}$ , printed from bottom to top, and a sequence of stack operations drawn from the set  $\{[\mathrm{PUSH}\mathsf{a}],[\mathrm{PUSH}\mathsf{b}],[\mathrm{POP}]\}$ . The function  $f_{\mathrm{SM}}(\pmb{x})$  outputs the final stack after all the given operations are executed sequentially on the input stack, printed from top to bottom. We always have  $|\pmb{y}| = |\pmb{x}| + 1$ . If the final stack has fewer elements than  $|\pmb{y}|$ , it will be padded with [PAD] tokens, which are ignored when accuracy is computed. We have  $\Sigma_{\mathrm{I}} = \{\mathsf{a},\mathsf{b},[\mathrm{PUSH}\mathsf{a}],[\mathrm{PUSH}\mathsf{b}],[\mathrm{POP}]\}$  and
+
+$\Sigma_{\mathrm{O}} = \{\mathsf{a},\mathsf{b},[\mathrm{PAD}]\}$ . An example is given below.
+
+# Example:
+
+$$
+\boldsymbol {x} = \operatorname {b a b} [ \mathrm {P O P} ] [ \mathrm {P U S H} \mathrm {a} ] [ \mathrm {P U S H} \mathrm {b} ]
+$$
+
+$$
+\boldsymbol {y} = \text {b a a b} [ \mathrm {P A D} ] [ \mathrm {P A D} ] [ \mathrm {P A D} ]
+$$
+
+Modular Arithmetic (MA). In the third task, we consider a transduction task based on modular arithmetic. An algebraic expression  $\pmb{x}$  consists of five numerical constants  $\{0,1,2,3,4\}$ , three operations  $\{+, -, \cdot\}$ , brackets  $\{(.,)\}$ , and a congruence sign  $\{\equiv\}$ . We say two integers are congruent if and only if a pre-set modulus is a divisor of their difference. In this task, we set the modulus to 5, so the function  $f_{\mathrm{MA}}$  evaluates the expression modulo 5. We have  $\Sigma_{\mathrm{I}} = \{0,1,2,3,4,+, -, \cdot, (\cdot), \equiv\}$  and  $\Sigma_{\mathrm{O}} = \{0,1,2,3,4\}$ . An example is given below.
+
+# Example:
+
+$$
+\boldsymbol {x} = (1 + 2) \cdot 3 \equiv
+$$
+
+$$
+\boldsymbol {y} = 4
+$$
+
+Solve Equation (SE). In our fourth and final task, we consider a transduction task that solves equations over a single variable, which we denote  $z$ . The input string  $x$  is a modular equation with five constants  $\{0,1,2,3,4\}$ , two operations  $\{+, - \}$ , brackets  $\{(.,)\}$ , a congruence sign  $\{\equiv \}$ , and a variable  $\{z\}$ . The modulus is set to 5. The function  $f_{\mathrm{SE}}$  solves this equation and returns the value of the variable. We have  $\Sigma_{\mathrm{I}} = \{0,1,2,3,4, + , - ,(.,)\equiv ,z\}$  and  $\Sigma_{\mathrm{O}} = \{0,1,2,3,4\}$ . An example is given below.
+
+# Example:
+
+$$
+\boldsymbol {x} = (1 + z) + 2 \equiv 2
+$$
+
+$$
+\boldsymbol {y} = 4
+$$
+
+# 4.2 Experimental Setup
+
+Following Delétang et al. (2023), we experiment with a transformer with the number of layers  $L = 5$  and the model size  $D = 64$ . Unless otherwise specified, no positional encodings are used. We discuss the effect of various positional encodings in §4.3.2. Length generalization has been the focus of many papers in this line of research (Joulin and Mikolov, 2015; Delétang et al., 2023). We follow suit to train on input strings  $x$  with  $1 \leqslant |\pmb{x}| \leqslant 40$  and test on  $x$  with  $40 < |\pmb{x}| \leqslant 100$ . Training details can be found in Appendix B.
+
+# 4.3 Results
+
+We report our results of the four DCF tasks in Tab. 1 and Tab. 2.
+
+# 4.3.1 Transformer vs. Stack Transformer
+
+We report the performance of the standard transformer and our stack-augmented transformer on the four DCF tasks presented in Tab. 1. For comparison, we also exhibit results of vanilla recurrent neural networks (RNNs) and stack-RNNs (Joulin and Mikolov, 2015). As expected, the vanilla transformer exhibits poor performance on all the DCF tasks. After being augmented with a stack, the transformer improves from nearly chance to over  $90\%$  on RS and SM. These results demonstrate that our stack-augmented attention helps on some tasks. However, on MA and SE, the performance after adding the stack attention only improves slightly; it still falls far behind stack RNNs and even vanilla RNNs. We conjecture that the reason for this shortcoming is our stack transformer's incapability to learn non-counter-free languages—both the
+
+last two tasks (MA) and SE require the ability to perform modular arithmetic, which makes them non-star-free, as discussed in §3.6. Additionally, Feng et al. (2023) also directly prove that the transformer cannot perform modular arithmetic.
+
+# 4.3.2 Positional Encodings
+
+In this section, we add various positional encodings to the transformer and investigate their effect. We consider five different positional encodings: none, sincos, relative, rotary, and ALiBi; see Appendix C for more details. As our stack attention is computed inductively, positional information is already present in the model, reducing the need for positional encodings. This is evident in Tab. 2, where including positional encodings generally has a negative impact on the stack transformer's performance. Most notably, sincos and rotary heavily degrade stack transformer's performance on RS and SM. However, relative constitutes an exception, as it results in improved performance on SM. In contrast, with the standard transformer architecture, positional encodings do seem to help on star-free tasks. The largest improvement comes from ALiBi in the MLM setting and rotary in the ALM setting. Nevertheless, none of the investigated positional encodings are able to boost the performance of vanilla transformer to anywhere near that of our stack-augmented transformer.
+
+# 5 Language Modeling
+
+We consider masked language modeling using RoBERTa (Liu et al., 2020) and autoregressive language modeling using GPT-2 (Radford et al., 2019). Following the experimental setup proposed by previous authors (Joulin and Mikolov, 2015; DuSell and Chiang, 2024), we experiment on the Penn Treebank (PTB), licensed through the LDC (Marcus et al., 1994), and WikiText-2 (Merit et al., 2017). We consider models both trained from scratch and fine-tuned from pre-trained weights. The pre-trained models and datasets are obtained from HuggingFace (Wolf et al., 2020; Lhoest et al., 2021). See Appendix B for more details about setup and hyperparameters.
+
+The results in Tab. 3 are mixed. Our major finding is that transformers benefit from the stack attention when training data is scarce, but the benefits gradually diminish as the size of training data grows. More concretely, when the models are trained from scratch, the addition of our stack attention mechanism does result in a noticeable benefit
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">Task</td><td colspan="2">Penn Treebank</td><td colspan="2">WikiText-2</td></tr><tr><td>Vanilla</td><td>Stack</td><td>Vanilla</td><td>Stack</td></tr><tr><td rowspan="2">Scratch</td><td>MLM</td><td>95.53 ± 19.66</td><td>34.28 ± 2.76</td><td>73.74 ± 3.79</td><td>64.75 ± 1.75</td></tr><tr><td>ALM</td><td>73.14 ± 0.34</td><td>69.86 ± 0.26</td><td>191.01 ± 0.71</td><td>206.42 ± 0.80</td></tr><tr><td rowspan="2">Pretrained</td><td>MLM</td><td>3.99 ± 0.08</td><td>4.46 ± 0.11</td><td>4.41 ± 0.12</td><td>4.65 ± 0.06</td></tr><tr><td>ALM</td><td>21.26 ± 0.03</td><td>32.36 ± 0.16</td><td>29.29 ± 0.02</td><td>54.96 ± 0.19</td></tr></table>
+
+Table 3: MLM and ALM Perplexities on WikiText-2 and PTB.
+
+in most settings. In the MLM setting, where  $15\%$  of the tokens are replaced with [MASK], stacks reduce the perplexity under the trained model on the held-out split from 95.53 to 34.28 on PTB and from 73.74 to 65.22 on WikiText-2. In the ALM setting, the stack transformer still slightly improves the performance on PTB—perplexity drops from 73.14 to 69.86. However, the stack transformer is less effective on WikiText-2, whose training set is larger. Moreover, when we fine-tune from pre-trained models, stacks are always detrimental across the two datasets in both MLM and ALM settings.
+
+# 6 Discussion
+
+From the results described in  $\S 4.3$  and  $\S 5$ , we observe two trends:
+
+- The positive impact of stack attention is evident on Delétang et al.'s (2023) 4 DCF tasks (especially on 2 of the 4), but almost nonexistent on English language modeling;  
+- On the English language modeling task, stack attention is more helpful in settings with limited training data, but is less helpful and can even be harmful when the model is trained on a larger amount of data.
+
+We interpret these trends as support for the idea that stack attention improves the representational capacity of a transformer language model and, additionally, confers an inductive bias to the transformer architecture that allows it to better learn certain context-free tasks more efficiently. The larger representational capacity explains why the performance on certain tasks, i.e., RS and SM, improves drastically with the addition of stack attention and the better inductive bias explains why transformer language models with stack attention perform better with less training data on the English modeling task. However, the fact that a vanilla transformer language model performs on par with stack attention when modeling larger
+
+English language datasets suggests that a good inductive bias is not needed for larger data sets. This suggests that, in contrast to the viewpoint of traditional linguistic theory (Chomsky, 1957), models that are higher on the Chomsky hierarchy are not necessary for developing a good statistical language model. We believe this claim is consistent with the literature, in which few successful large language models are endowed with a syntactic bias. However, there are many smaller syntax-infused language models (Dyer et al., 2016) that do work well on smaller data, as ours does.
+
+# 7 Conclusion
+
+We propose a novel implementation of a differentiable stack and show that a transformer augmented with such stacks can solve certain deterministic context-free tasks that are beyond the capability of standard transformers. However, unlike a stack RNN, the stack transformer cannot model the entire class of deterministic context-free languages.
+
+# Acknowledgements
+
+This publication was made possible by an ETH AI Center doctoral fellowship to Jiaoda Li. Ryan Cotterell acknowledges support from the Swiss National Science Foundation (SNSF) as part of the "The Forgotten Role of Inductive Bias in Interpretability" project.
+
+# Limitations
+
+The primary limitation of the proposed stack attention is it only allows one POP operation at a time. It can be extended to have multiple POPs in a manner similar to Yogatama et al. (2018). It can also only handle deterministic context-free languages. We would like to extend it to non-deterministic stacks in future works. Although our method does not require structural supervision, it can in principle take advantage of it when it is available. In such cases,
+
+the model can be fully parallelized, leading to great improvement in time efficiency. It would be interesting to explore this possibility in the future.
+
+# Ethical Considerations
+
+We foresee no ethical concerns with this work.
+
+# References
+
+Dana Angluin, David Chiang, and Andy Yang. 2023. Masked hard-attention transformers and boolean RASP recognize exactly the star-free languages. Computing Research Repository, arXiv:2310.13897. Version 2.  
+Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. 2016. Layer normalization. Computing Research Repository, arXiv:1607.06450.  
+Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2015. Neural machine translation by jointly learning to align and translate. In International Conference on Learning Representations.  
+Pablo Barcelo, Alexander Kozachinskiy, Anthony Widjaja Lin, and Vladimir Podolskii. 2024. Logical languages accepted by transformer encoders with hard attention. In International Conference on Learning Representations.  
+Satwik Bhattachamishra, Kabir Ahuja, and Navin Goyal. 2020. On the Ability and Limitations of Transformers to Recognize Formal Languages. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7096-7116.  
+Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D. Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems, volume 33, pages 1877-1901.  
+Jianpeng Cheng, Li Dong, and Mirella Lapata. 2016. Long short-term memory-networks for machine reading. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 551-561, Austin, Texas. Association for Computational Linguistics.  
+David Chiang, Peter Cholak, and Anand Pillay. 2023. Tighter bounds on the expressivity of transformer encoders. In Proceedings of the 40th International Conference on Machine Learning, ICML'23. JMLR.org.  
+Noam Chomsky. 1956. Three models for the description of language. IRE Transactions on Information Theory, 2(3):113-124.
+
+Noam Chomsky. 1957. Syntactic Structures. De Gruyter Mouton, Berlin, Boston.  
+Noam Chomsky and Marcel Paul Schützenberger. 1963. The algebraic theory of context-free languages. Studies in logic and the foundations of mathematics, 35:118-161.  
+Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc Le, and Ruslan Salakhutdinov. 2019. Transformer-XL: Attentive language models beyond a fixed-length context. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2978-2988.  
+Sreerupa Das, C. Lee Giles, and Guo-Zheng Sun. 1992. Learning context-free grammars: Capabilities and limitations of a recurrent neural network with an external stack memory. In Proceedings of The Fourteenth Annual Conference of Cognitive Science Society, volume 14.  
+Grégoire Deletang, Anian Ruoss, Jordi Grau-Moya, Tim Genewein, Li Kevin Wenliang, Elliot Catt, Chris Cundy, Marcus Hutter, Shane Legg, Joel Veness, and Pedro A. Ortega. 2023. Neural networks and the Chomsky hierarchy. In 11th International Conference on Learning Representations.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186.  
+Brian DuSell and David Chiang. 2020. Learning context-free languages with nondeterministic stack RNNs. In Proceedings of the 24th Conference on Computational Natural Language Learning, pages 507-519.  
+Brian DuSell and David Chiang. 2022. Learning hierarchical structures with differentiable nondeterministic stacks. In International Conference on Learning Representations.  
+Brian DuSell and David Chiang. 2024. Stack attention: Improving the ability of transformers to model hierarchical patterns. In International Conference on Learning Representations.  
+Chris Dyer, Adhiguna Kuncoro, Miguel Ballesteros, and Noah A. Smith. 2016. Recurrent neural network grammars. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 199-209, San Diego, California. Association for Computational Linguistics.  
+Javid Ebrahimi, Dhruv Gelda, and Wei Zhang. 2020. How can self-attention networks recognize Dyck-n languages? In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 4301-4306.  
+R. James Evey. 1963. Application of pushdown-store machines. In Proceedings of the November 12-14, 1963, Fall Joint Computer Conference, AFIPS '63
+
+(Fall), page 215-227, New York, NY, USA. Association for Computing Machinery.  
+Guhao Feng, Bohang Zhang, Yuntian Gu, Haotian Ye, Di He, and Liwei Wang. 2023. Towards revealing the mystery behind chain of thought: A theoretical perspective. In Advances in Neural Information Processing Systems.  
+Edward Grefenstette, Karl Moritz Hermann, Mustafa Suleyman, and Phil Blunsom. 2015. Learning to transduce with unbounded memory. In Advances in Neural Information Processing Systems, volume 28.  
+Michael Hahn. 2020. Theoretical limitations of self-attention in neural sequence models. Transactions of the Association for Computational Linguistics, 8:156-171.  
+Yiding Hao, Dana Angluin, and Robert Frank. 2022. Formal language recognition by hard attention transformers: Perspectives from circuit complexity. Transactions of the Association for Computational Linguistics, 10:800-810.  
+Yiding Hao, William Merrill, Dana Angluin, Robert Frank, Noah Amsel, Andrew Benz, and Simon Mendelsohn. 2018. Context-free transductions with neural stacks. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 306-315.  
+Armand Joulin and Tomas Mikolov. 2015. Inferring algorithmic patterns with stack-augmented recurrent nets. In Advances in Neural Information Processing Systems, volume 28.  
+Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In International Conference on Learning Representations.  
+Quentin Lhoest, Albert Villanova del Moral, Yacine Jernite, Abhishek Thakur, Patrick von Platen, Suraj Patil, Julien Chaumont, Mariama Drame, Julien Plu, Lewis Tunstall, Joe Davison, Mario Šaško, Gunjan Chhablani, Bhavitvya Malik, Simon Brandeis, Teven Le Scao, Victor Sanh, Canwen Xu, Nicolas Patry, Angelina McMillan-Major, Philipp Schmid, Sylvain Gugger, Clément Delangue, Theo Matussière, Lysandre Debut, Stas Bekman, Pierric Cistac, Thibault Goehringer, Victor Mustar, François Lagunas, Alexander Rush, and Thomas Wolf. 2021. Datasets: A community library for natural language processing. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 175-184, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2020. RoBERTa: A robustly optimized BERT pretraining approach.  
+Mitchell Marcus, Grace Kim, Mary Ann Marcinkiewicz, Robert MacIntyre, Ann Bies, Mark Ferguson, Karen Katz, and Britta Schasberger. 1994. The Penn Treebank: Annotating predicate argument structure. In Human Language Technology: Proceedings of a
+
+Workshop held at Plainsboro, New Jersey, March 8-11, 1994.  
+Robert McNaughton and Seymour Papert. 1971. Counter-free automata. Research monograph 65. MIT Press.  
+Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. 2017. Pointer sentinel mixture models. In International Conference on Learning Representations.  
+William Merrill and Ashish Sabharwal. 2023. A logic for expressing log-precision transformers. In Advances in Neural Information Processing Systems, volume 36, pages 52453–52463. Curran Associates, Inc.  
+William Merrill and Ashish Sabharwal. 2024. The expressive power of transformers with chain of thought. In International Conference on Learning Representations.  
+Michael C. Mozer and Sreerupa Das. 1992. A connectionist symbol manipulator that discovers the structure of context-free languages. In Advances in Neural Information Processing Systems, volume 5, pages 863-870.  
+Shikhar Murty, Pratyusha Sharma, Jacob Andreas, and Christopher Manning. 2023. Pushdown layers: Encoding recursive structure in transformer language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 3233-3247, Singapore. Association for Computational Linguistics.  
+Anthony G. Oettinger. 1961. Automatic syntactic analysis and the pushdown store. In Proceedings of Symposia in Applied Mathematics, volume 12: Structure of Language and Its Mathematical Aspects, pages 104-129, Providence, Rhode Island.  
+Ankur Parikh, Oscar Täckström, Dipanjan Das, and Jakob Uszkoreit. 2016. A decomposable attention model for natural language inference. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2249-2255, Austin, Texas. Association for Computational Linguistics.  
+Jorge Pérez, Pablo Barceló, and Javier Marinkovic. 2021. Attention is Turing-complete. Journal of Machine Learning Research, 22(75):1-35.  
+Jordan B. Pollack. 1991. The induction of dynamical recognizers. Machine Learning, 7(2-3):227-252.  
+Ofir Press, Noah Smith, and Mike Lewis. 2022. Train short, test long: Attention with linear biases enables input length extrapolation. In International Conference on Learning Representations.  
+Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners.  
+Laurent Sartran, Samuel Barrett, Adhiguna Kuncoro, Miloš Stanojevic, Phil Blunsom, and Chris Dyer. 2022. Transformer grammars: Augmenting transformer language models with syntactic inductive biases at scale. Transactions of the Association for Computational Linguistics, 10:1423-1439.
+
+Marcel Paul Schützenberger. 1963. On context-free languages and push-down automata. Information and Control, 6(3):246-264.  
+M.P. Schützenberger. 1965. On finite monoids having only trivial subgroups. Information and Control, 8(2):190-194.  
+Stuart M. Shieber. 1987. Evidence Against the Context-Freeness of Natural Language, pages 320-334. Springer Netherlands.  
+Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. 2023. RoFormer: Enhanced transformer with rotary position embedding. Neurocomputing, page 127063.  
+Mirac Suzgun, Sebastian Gehrmann, Yonatan Belinkov, and Stuart M. Shieber. 2019. Memory-augmented recurrent neural networks can learn generalized Dyck languages. Computing Research Repository, arXiv:1911.03329.  
+Anej Svete and Ryan Cotterell. 2024. Transformers can represent  $n$ -gram language models. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics, Mexico City, Mexico. Association for Computational Linguistics.  
+The Univalent Foundations Program. 2013. Homotopy Type Theory: Univalent Foundations of Mathematics. Institute for Advanced Study.  
+Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30.  
+Gail Weiss, Yoav Goldberg, and Eran Yahav. 2021. Thinking like transformers. In Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 11080-11090. PMLR.  
+Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45.  
+Dani Yogatama, Yishu Miao, Gabor Melis, Wang Ling, Adhiguna Kuncoro, Chris Dyer, and Phil Blunsom. 2018. Memory architectures in recurrent neural network language models. In International Conference on Learning Representations.  
+Zheng Zeng, Rodney M. Goodman, and Padhraic Smyth. 1994. Discrete recurrent neural networks for grammatical inference. IEEE Transactions on Neural Networks, 5(2):320-330.
+
+# A Proof
+
+Theorem 3.1. Let  $v_{1}, \ldots, v_{N}$  be a series of stack operations where  $v_{i} \in \{\mathsf{PUSH}_{i}(\cdot), \mathsf{NO - OP}(\cdot), \mathsf{POP}(\cdot)\}$  for all  $i \in [N]$ . Furthermore, suppose  $\mathbf{a}_{i}(v_{i}) = 1$  for all  $i \in [N]$ . Then,  $[[\mathsf{PEEK}(v_{i}(\dots v_{1}(\varepsilon)))]] = \alpha_{i}$  for all  $i \in \{0\} \cup [N]$ .
+
+Proof.
+
+Base case  $(i = 0)$ . The stack is initialized to be empty, i.e.,  $\gamma_0 = \varepsilon$  and  $\mathsf{PEEK}(\gamma_0) = 0$ . By definition, we have
+
+$$
+\boldsymbol {\alpha} _ {0} = [ 1, 0, \dots ] ^ {\top} = \llbracket \operatorname {P E E K} (\boldsymbol {\gamma} _ {0}) \rrbracket \tag {28}
+$$
+
+Inductive Step. Suppose there exists an  $i > 0$ , such that  $\forall i' < i$ ,  $\alpha_{i'} = \mathbb{[PEEK}(\gamma_{i'})\mathbb{]}$ , and  $\mathbf{a}_i(v_i) = 1$ .
+
+- If  $v_{i} = \mathsf{PUSH}, \gamma_{i} = \mathsf{PUSH}(\gamma_{i - 1}) = \gamma_{i - 1}i$  and  $\mathsf{PEEK}(\gamma_i) = i$ , so according to Eq. (20) we have  $\alpha_{i} = [[\mathsf{PEEK}(\gamma_{i})]]$ .  
+- If  $v_{i} = \mathsf{NO - OP}$ ,  $\gamma_{i} = \mathsf{NO - OP}(\gamma_{i - 1}) = \gamma_{i - 1}$ , and  $\alpha_{i} = \alpha_{i - 1}$ . Since  $\alpha_{i - 1} = \llbracket \mathsf{PEEK}(\gamma_{i - 1})\rrbracket$ , we have  $\alpha_{i} = \llbracket \mathsf{PEEK}(\gamma_{i})\rrbracket$ .  
+- If  $v_{i} = \mathsf{POP}$
+
+$$
+\boldsymbol {\alpha} _ {i} = \sum_ {j = 1} ^ {i - 1} \boldsymbol {\alpha} _ {i - 1} (j) \boldsymbol {\alpha} _ {j - 1} + \boldsymbol {\alpha} _ {i - 1} (0) \boldsymbol {\alpha} _ {0} \tag {29}
+$$
+
+If  $\alpha_{i - 1}(0) = 1$ , i.e.  $\gamma_{i - 1}$  is empty,  $\gamma_{i} = \mathsf{POP}(\varepsilon) = \varepsilon$ , and
+
+$$
+\begin{array}{l} \boldsymbol {\alpha} _ {i} = \boldsymbol {\alpha} _ {i - 1} (0) \boldsymbol {\alpha} _ {0} (30a) \\ = \alpha_ {0} (30b) \\ = [ 0 ] (30c) \\ = \llbracket \operatorname {P E E K} \left(\gamma_ {i}\right) \rrbracket (30d) \\ \end{array}
+$$
+
+Otherwise,
+
+$$
+\begin{array}{l} \boldsymbol {\alpha} _ {i} = \sum_ {j = 1} ^ {i - 1} \boldsymbol {\alpha} _ {i - 1} (j) \boldsymbol {\alpha} _ {j - 1} (31a) \\ = \alpha_ {\text {P E E K} \left(\gamma_ {i - 1}\right) - 1} (31b) \\ = \llbracket \text {P E E K} \left(\gamma_ {\text {P E E K}} \left(\gamma_ {i - 1}\right) - 1\right) \rrbracket (31c) \\ = \llbracket \mathrm {P E E K} (\mathrm {P O P} (\gamma_ {i - 1})) \rrbracket (31d) \\ = \llbracket \mathrm {P E E K} (\boldsymbol {\gamma} _ {i}) \rrbracket (31e) \\ \end{array}
+$$
+
+One can understand Eq. (31d) intuitively as follows:  $\gamma_{\mathrm{PEEK}(\gamma_{i - 1}) - 1}$  is the stack right before the current stack top  $\mathsf{PEEK}(\gamma_{i - 1})$  is pushed, so the stack top at  $\mathsf{PEEK}(\gamma_{i - 1}) - 1$  is the second top-most element at  $i - 1$ , i.e.,  $\mathsf{PEEK}(\gamma_{\mathsf{PEEK}(\gamma_{i - 1}) - 1}) = \mathsf{POP}(\gamma_{i - 1})$ .
+
+Theorem 3.2. Consider a sequence of stack attention mechanisms  $\alpha_0, \ldots, \alpha_N$ . Then,  $\sum_{n=0}^{N} \alpha_i(n) = 1$  for all  $i \in \{0\} \cup [N]$ .
+
+Proof.
+
+Base case It holds for  $\alpha_0 = [1,0,\dots ]^\top$
+
+Induction step Suppose there exists an  $i > 0$ , such that  $\forall i' < i, \sum_{n=0}^{N} \alpha_{i'}(n) = 1$ .
+
+- PUSH. Obviously,
+
+$$
+\sum_ {n = 0} ^ {N} \boldsymbol {\alpha} _ {i} ^ {(\text {P U S H})} (n) = 1 \tag {32}
+$$
+
+- NO-OP. Since  $\alpha_{i}^{(\mathrm{NO - OP})} = \alpha_{i - 1}$ , we also have
+
+$$
+\sum_ {n = 0} ^ {N} \alpha_ {i} ^ {\left(\mathrm {N O} - \mathrm {O P}\right)} (n) = 1 \tag {33}
+$$
+
+POP.
+
+$$
+\begin{array}{l} \sum_ {n = 0} ^ {N} \boldsymbol {\alpha} _ {i} ^ {(\mathrm {P O P})} (j) = \sum_ {n = 0} ^ {N} \left(\sum_ {j = 1} ^ {i - 1} \boldsymbol {\alpha} _ {i - 1} (j) \boldsymbol {\alpha} _ {j - 1} + \boldsymbol {\alpha} _ {i - 1} (0) \boldsymbol {\alpha} _ {0}\right) (n) (34a) \\ = \sum_ {n = 0} ^ {N} \left(\sum_ {j = 1} ^ {i - 1} \boldsymbol {\alpha} _ {i - 1} (j) \boldsymbol {\alpha} _ {j - 1} (n) + \boldsymbol {\alpha} _ {i - 1} (0) \boldsymbol {\alpha} _ {0} (n)\right) (34b) \\ = \sum_ {j = 1} ^ {i - 1} \boldsymbol {\alpha} _ {i - 1} (j) \left(\sum_ {n = 0} ^ {N} \boldsymbol {\alpha} _ {j - 1} (n)\right) + \boldsymbol {\alpha} _ {i - 1} (0) \left(\sum_ {n = 0} ^ {N} \boldsymbol {\alpha} _ {0} (n)\right) (34c) \\ = \sum_ {j = 1} ^ {i - 1} \boldsymbol {\alpha} _ {i - 1} (j) + \boldsymbol {\alpha} _ {i - 1} (0) (34d) \\ = \sum_ {j = 0} ^ {i - 1} \alpha_ {i - 1} (j) (34e) \\ = 1 (34f) \\ \end{array}
+$$
+
+Therefore,
+
+$$
+\begin{array}{l} \sum_ {n = 0} ^ {N} \alpha_ {i} (n) (35a) \\ = \sum_ {n = 0} ^ {N} \left(\sum_ {a \in \mathcal {A}} \mathbf {a} _ {i} (a) \boldsymbol {\alpha} _ {i} ^ {(a)}\right) (n) (35b) \\ = \sum_ {a \in \mathcal {A}} \mathbf {a} _ {i} (a) \sum_ {n = 0} ^ {N} \boldsymbol {\alpha} _ {i} ^ {(a)} (n) (35c) \\ = \sum_ {a \in \mathcal {A}} \mathbf {a} _ {i} (a) (35d) \\ = 1 (35e) \\ \end{array}
+$$
+
+# B Experimental Setup
+
+# B.1 DCF Tasks
+
+The model is trained using the Adam optimizer (Kingma and Ba, 2015) with a learning rate of  $1e^{-4}$ , which we find works well for all the tasks. On the RS and SM tasks, we use a batch size of 32 and we train the model for 100,000 steps. On the MA and SE tasks, the batch size and the number of training steps are increased to 128 and 1,000,000, respectively, to ensure sufficient training. Each experiment is run 5 times with different random seeds. Means and variances of accuracies are reported Tab. 1 and Tab. 2.
+
+# B.2 Language Modeling
+
+The texts in the datasets are grouped into chunks of 128 tokens. Each model is, again, trained using the Adam optimizer for a maximum of 100 epochs with early stopping applied when the validation loss has not improved for 5 epochs in a row. We tune the learning rate from  $\{1e^{-5}, 2e^{-5}, 1e^{-4}, 2e^{-4}\}$  on the validation set, and choose  $2e^{-5}$  that leads to the best validation performance. Results on the test set over 5 runs with different random seeds are reported in Tab. 3. Experiments are conducted on a single NVIDIA Tesla V100 GPU.
+
+# C Positional Encodings
+
+We consider five different commonly used positional encodings:
+
+- none. No positional encodings are used.  
+- sincos. The sinusoidal positional encodings used in vanilla transformer (Vaswani et al., 2017). Positional information encoded sinusoidally is added to the embeddings.  
+- relative. In Transformer-XL (Dai et al., 2019), relative rather than absolute sinusoidal positional information is added to the keys and queries of each attention block.  
+- rotary. Introduced by Su et al. (2023) and popularized by GPT-3 (Brown et al., 2020), rotary positional encodings multiply the keys and queries by sinusoidal encodings.  
+- ALiBi. Press et al. (2022) adds linear biases to the attention blocks that favor the more recent tokens
+
+# D Analysis: Attention Maps
+
+An advantage of our stack attention mechanism is that we can visualize the stack tops  $\alpha_{i}$ , which provides greater interpretability than methods where stack tops are mixtures of hidden states (Joulin and Mikolov, 2015). We run a set of toy experiments with the stack transformer in the MLM setting. We randomly select one test example for each task.
+
+RS. At the first two layers (Fig. 2a, Fig. 2b), the first 5 tokens attend to themselves while the [MASK] tokens attend to the last token in  $x$ . The most probable sequence of operations that leads to such a stack attention map is the input  $x$  is pushed one by one onto the stack and NO-OP is performed on all the [MASK] tokens. At the third layer (Fig. 2c), the stacks for the [MASK] tokens shift one position backward at a time, which demonstrates the stack elements are POPED one by one to generate the outputs. At the last two layers, all the tokens attend to themselves, so the stacks can be regarded as being skipped (Fig. 2d, Fig. 2e).
+
+SM. Looking at the attention map at the first layer (Fig. 3a), we can infer the operations taken by the stack as follows: the stack first PUSHes the initial stack contents (ab); once the [POP] operation is read, it reverts to the first element a; then it performs the operation [PUSH b] twice as instructed; afterwards, it POPs the final stack bba for outputs. The stack attention correctly skips the b at timestep 1 as it has already been POPed at timestep 2. The last three positions are [PAD] tokens and can be ignored.
+
+MA and SE. We also provide an attention map for MA and SE in Fig. 4 and Fig. 5. Their attention maps are less interpretable as the stack transformer does not learn them well. Nevertheless, we can still observe that the stacks seem to be able to match the parentheses, which matches our expectations of the stack's strengths. For MA, at the first layer (Fig. 4a), the stack successfully matches the last two closing parentheses (at timestep 8 and 9) with their corresponding open parentheses (at timestep 5 and 0 respectively). For SE, the pattern is less obvious presumably because the parentheses do not have an impact on the order of arithmetic operations and can be ignored.
+
+![](images/31e377c7f2fb4c35bd64b6fdd424e2b8b81169ba756ab33013ca05cc4cc529d5.jpg)  
+(a) Layer 1
+
+![](images/7562b05a1f1c8b94fb6ae6b9424a212af28cf77b6e8012c4a13ce074e20587ff.jpg)  
+(b) Layer 2
+
+![](images/7f21767622039e8ae8cf83a858cc04ea20baf5b56ea304eb0afa710cdd37b293.jpg)
+
+![](images/2a44c7c248ee621cac841bc0760bac81074f630b24ac46cbf1afa772b55b2d56.jpg)  
+(d) Layer 4
+
+![](images/c1a8d7d550fa4d42aab060385c54d5ba322b2b3539e93f9fefc6adbe8a70deca.jpg)  
+(c) Layer 3  
+(e) Layer 5
+
+![](images/28a702da5ff050e32ad72ff143634f3d6e42de647334baadcd4cd0d74dabc185.jpg)  
+Figure 2: Stack attention maps at different layers for RS. The input  $x$  is abbaa. M represents a [MASK] token.
+
+![](images/e86029785efc76bdce99ee3f8f3ba6395cd4b7bd6b3a547922f98eb6d66850d0.jpg)  
+(b) Layer 2
+
+![](images/4e6f252706152e3b8a1e2ec99ee602959fe849203a072bf31d4134e645e16b48.jpg)
+
+![](images/b5dfc0cad1dcc65d085c70b2698fff8a82e4ca3931d126e61a2d3c6946107c09.jpg)  
+(a) Layer 1  
+(d) Layer 4  
+Figure 3: Stack attention maps at different layers for SM. The input  $x$  is ab[POP][PUSH a][PUSH b]. In the graphs, [PUSH a], [PUSH b], and [POP] are abbreviated as a, b, and P respectively. M represents a [MASK] token. The correct output should be bba followed by [PAD] tokens.
+
+![](images/691b5bd9ec04949707fae8be1379af2f63cc8d74e0d4426518c2d2eaf50dee5c.jpg)  
+(c) Layer 3  
+(e) Layer 5
+
+![](images/f52beb5c9c802307dc3bc93d8e6a6c522c989131b2a02d1bdc6ad58e89491dc2.jpg)  
+(a) Layer 1
+
+![](images/dae9342f1f233609ac067c42ead2c8ec2eb6b02123def33348f836b84c661ded.jpg)  
+(b) Layer 2
+
+![](images/5efcf28df1ce47da27bf480b3a1ba20f6b57cc0d50864f590ed22ab473a5f74a.jpg)
+
+![](images/ebcb9de9f899c9bf655d1a15858b93db8779b8a541ffc006a595830154206373.jpg)  
+(d) Layer 4
+
+![](images/7b87ebe0f27747cffdec7e891dd948590cd2c97a4bea9d0d033f6e48f40e1358.jpg)  
+(c) Layer 3  
+(e) Layer 5
+
+![](images/477ebd1b81babce9a436e0dc3bc6049abb9a9e8e67355db653dd2be5c1b27d32.jpg)  
+Figure 4: Stack attention maps at different layers for MA. The input  $\pmb{x}$  is  $((4)\cdot (-0)) =$ .
+
+![](images/7637919383125a34c4c3da366412309ac9270111f001d074b40988b48c6f61b0.jpg)  
+(b) Layer 2
+
+![](images/4367b4252bf05a2f24eae6ab3bb9972cc1069294d3e6d027a9fc771d96bfbba8.jpg)
+
+![](images/86364815094f1fda5ea3aededf6951367a8588e4af24e1cfc56aef57e1c8ce88.jpg)  
+(a) Layer 1  
+(d) Layer 4  
+Figure 5: Stack attention maps at different layers for SE. The input  $\pmb{x}$  is  $(1 + (-z)) = 3$ . M represents a [MASK] token.
+
+![](images/b2303a43d8f2bdae68edb4c21def8c27b1a1aed4721152f6dc912916f91d3b65.jpg)  
+(c) Layer 3  
+(e) Layer 5
+
+# E Related Work
+
+# E.1 Stack Augmentation
+
+Equipping a neural network with a data structure such as an external stack to enhance its ability to recognize context-free languages has been extensively investigated in previous works (Pollack, 1991; Das et al., 1992; Mozer and Das, 1992; Zeng et al., 1994). The idea has seen a resurgence in recent years, with work focusing primarily on recurrent networks (Joulin and Mikolov, 2015; Grefenstette et al., 2015; Hao et al., 2018; Yogatama et al., 2018; Suzgun et al., 2019; DuSell and Chiang, 2020, 2022). Joulin and Mikolov (2015) propose to superpose the result of applying each stack operation at each step, which directly inspires our work. We adapt it for application to transformers by rendering this concept as an attention mechanism. In that sense, our work is related to Das et al. (1992) and Grefenstette et al. (2015), which also assign weights to stack elements. Our stack attention mechanism is different as the stack attention weights are assigned to previously seen tokens indicating where the top element is located
+
+Sartran et al. (2022) and Murty et al. (2023) incorporate a stack mechanism into a transformer language model with structural supervision during training. DuSell and Chiang's (2024) contemporaneous work also augments a transformer language model with a stack. Both their and our methods are named stack attention, but their stack attention is an attention mechanism over stack actions while ours is an attention mechanism over input tokens.
+
+# E.2 Expressivity of Transformers
+
+The expressivity of transformers under various assumptions has been extensively studied. A stream of research considers transformer encoders with a classification layer at the end as recognizers. Hahn (2020) proves that transformers cannot recognize parity language, a periodic language of binary strings with an even number of 1's, and Dyck-2 language, a CF language of balanced brackets of two types. Bhattachamishra et al. (2020) find that transformers can recognize certain counter languages but fail to recognize non-star-free languages such as  $(\mathsf{aa})^{*}$ . Svete and Cotterell (2024) show that transformers can represent  $n$ -gram language models. Hao et al. (2022), Chiang et al. (2023), Merrill and Sabharwal (2023), Barcelo et al. (2024), and Angluin et al. (2023) relate transformers to circuit complexity and formal logic. With various extensions, transformers' expressivity can be increased. Weiss et al. (2021) propose a programming language that shares the same basic operations with transformers but is more expressive than standard transformers. Pérez et al. (2021) and Merrill and Sabharwal (2024) show that transformer encoder-decoders and decoders are Turing complete with additional scratch space.

2024/A Transformer with Stack Attention/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:db51312c909222daa352cbccd0077548bfefc6fe9b0438b58e6545d73a670ad2
+size 780548

2024/A Tree-of-Thoughts to Broaden Multi-step Reasoning across Languages/cf769ade-bf6e-4081-af4e-c7800f171438_content_list.json

@@ -0,0 +1,2143 @@
+[
+    {
+        "type": "text",
+        "text": "A Tree-of-Thoughts to Broaden Multi-step Reasoning across Languages",
+        "text_level": 1,
+        "bbox": [
+            127,
+            90,
+            870,
+            111
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Leonardo Ranaldi $(\\diamond)$ , Giulia Pucci  $(\\star)$ , Federico Ranaldi  $(\\diamond)$ , Elena Sofia Ruzzetti  $(\\diamond)$ , Fabio Massimo Zanzotto $(\\diamond)$",
+        "bbox": [
+            188,
+            127,
+            811,
+            161
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$(\\diamond)$  Human-Centric ART Group,",
+        "bbox": [
+            371,
+            162,
+            630,
+            178
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Dep. of Enterprise Engineering, University of Rome Tor Vergata, Italy",
+        "bbox": [
+            216,
+            179,
+            784,
+            195
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$(\\star)$ Department of Computing Science, University of Aberdeen, UK",
+        "bbox": [
+            228,
+            196,
+            773,
+            211
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "[first_name].[last_name]@uniroma2.it",
+        "bbox": [
+            319,
+            212,
+            682,
+            227
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            261,
+            339,
+            275
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Reasoning methods, best exemplified by the well-known Chain-of-Thought (CoT), empower the reasoning abilities of Large Language Models (LLMs) by eliciting them to solve complex tasks in a step-by-step manner. Although they are achieving significant success, the ability to deliver multi-step reasoning remains limited to English because of the imbalance in the distribution of pre-training data, which makes other languages a barrier.",
+        "bbox": [
+            141,
+            284,
+            460,
+            426
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we propose Cross-lingual Tree-of-Thoughts (Cross-ToT), a method for aligning Cross-lingual CoT reasoning across languages. The proposed method, through a self-consistent cross-lingual prompting mechanism inspired by the Tree-of-Thoughts approach, provides multi-step reasoning paths in different languages that, during the steps, lead to the final solution. Experimental evaluations show that our method significantly outperforms existing prompting methods by reducing the number of interactions and achieving state-of-the-art performance.",
+        "bbox": [
+            141,
+            426,
+            460,
+            598
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            607,
+            258,
+            621
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Chain-of-Thought (CoT) prompting elicits Large Language Models (LLMs) to break down a reasoning task towards a sequence of intermediate steps (Wei et al., 2022). Previous works have demonstrated that LLMs achieve impressive performances in zero-shot learning scenarios without the need to modify the model parameters during the training and testing process. In particular, by appending to the prompt \"Let's think step by step!\" (Kojima et al., 2023) LLMs with at least several billions of parameters, such as GPTs family (OpenAI, 2023) or PaLM (Chowdhery et al., 2022), deliver multi-step controlled reasoning, achieving promising results across commonsense (Bubeck et al., 2023), symbolic and mathematical reasoning datasets (Gaur and Saunshi, 2023; Liu et al., 2023).",
+        "bbox": [
+            112,
+            631,
+            489,
+            888
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Although the performances seem promising, they are only firmly established in English. This",
+        "bbox": [
+            112,
+            890,
+            487,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "poses a barrier to generalizing current CoT techniques to different languages. Hence, despite the remarkable success of zero-shot CoT techniques, the reasoning abilities of LLMs still struggle to generalize to different languages. Shi et al. (2022) introduced the first multilingual benchmark to assess LLMs' mathematical reasoning abilities using prompts in different languages. Qin et al. (2023) propose task-specific solver prompting, using a succession of prompts, elicit the LLMs to understand questions and deliver CoT answers in different languages. However, these strategies require two-step prompts, which goes against the zero-shot approach.",
+        "bbox": [
+            507,
+            261,
+            884,
+            485
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we propose Cross-lingual Tree-of-Thoughts (Cross-ToT), a method for aligning Cross-lingual CoT reasoning across languages by proposing a Cross-lingual Alignment prompt to elicit the model to deliver a Self-consistent Chain-of-Thought. Our method is inspired by the Tree-of-Thoughts (ToT) prompting (Yao et al., 2023) that asks LLMs to perform decision-making by considering multiple different reasoning paths (CoTs). In particular, our Cross-ToT is a ToT-style prompting to deliver the reasoning process in different languages that, step-by-step, converge to a single final solution. The inherent insight is that as the different paths of thought evolve, the relationships between the different languages are inherently grasped via Self-consistent Chains-of-Thought. This leads to the target research questions, which are the focus of this paper:",
+        "bbox": [
+            507,
+            487,
+            884,
+            774
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "RQ1: Are LLMs able to deliver Cross-lingual multi-step reasoned answers?",
+        "bbox": [
+            507,
+            776,
+            882,
+            807
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$RQ2$ : Are the different paths of ToT evolving Self-correcting each other?",
+        "bbox": [
+            507,
+            809,
+            880,
+            840
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$RQ3$ : What is the role of English in Cross-lingual scenarios?",
+        "bbox": [
+            507,
+            841,
+            882,
+            871
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To answer these questions, we propose Cross-ToT, a novel Cross-lingual prompting strategy that aims to bridge the gap across different",
+        "bbox": [
+            507,
+            873,
+            882,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "1229",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: NAACL 2024, pages 1229-1241 June 16-21, 2024 ©2024 Association for Computational Linguistics",
+        "bbox": [
+            218,
+            945,
+            778,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/ea195486126cd5fbe63d99bb4827d5b88b7fdb2929b011908c4c890d1eefd1ce.jpg",
+        "image_caption": [
+            "Figure 1: Our Cross-ToT elicits the LLM to generate step-by-step Cross-lingual reasoning. Furthermore, different pathways are developed during these reasoning steps. This mechanism develops the Chains-of-Thoughts in a Self-consistent way, streaming with the different pathways."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            139,
+            80,
+            884,
+            397
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "languages. In particular, using the prompt shown in Figure 1, we elicit the model to deliver different CoT reasoning steps in different languages that converge to the final solution step-by-step. We test our method on GPT-3.5 and conduct an extensive analysis using Multilingual Grade School Math (MGSM) (Shi et al., 2022), Cross-lingual Natural Language Inference (XNLI) (Conneau et al., 2018), and Cross-lingual Paraphrase Adversaries Scrambling (PAWS-X) (Yang et al., 2019), Cross-lingual Choice of Plausible Alternatives (XCOPA) (Ponti et al., 2020) across different languages. Experimental results reveal that our method, based on a single prompt, outperforms the baselines and achieves the SOTA performance on different languages in different tasks. The main contributions of this work are concluded as follows:",
+        "bbox": [
+            112,
+            480,
+            489,
+            752
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We introduce Cross-ToT, which is a novel Cross-lingual prompting mechanism that stimulates the model to produce parallel CoT reasoning processes across different languages;",
+            "- We show that our Cross-ToT is Self-consistent and allows the integration of reasoning paths between different languages;",
+            "- Extensive evaluations on different languages"
+        ],
+        "bbox": [
+            136,
+            766,
+            489,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "demonstrate that our Cross-ToT can effectively improve the performance of crosslingual CoTs and achieve SOTA performance.",
+        "bbox": [
+            544,
+            480,
+            884,
+            527
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "- Finally, we show that introducing English in our prompting technique plays a beneficial role in improving downstream performance.",
+        "bbox": [
+            531,
+            537,
+            880,
+            586
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Cross-lingual Multi-step Reasoning",
+        "text_level": 1,
+        "bbox": [
+            507,
+            593,
+            848,
+            609
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To elicit the multi-step reasoning abilities of LLMs in Cross-lingual scenarios, we propose Cross-ToT, which is a Cross-lingual Alignment Chain-of-Thought as a solution. In particular, our method overcomes the Multi-lingual and Cross-lingual approaches introduced in Section 2.1. In fact, our approach elicits the LLMs to deliver Self-consistent Parallel Chain-of-Thoughts, introduced in Section 2.2.",
+        "bbox": [
+            507,
+            617,
+            884,
+            760
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1 Chain-of-Thought Across Languages",
+        "text_level": 1,
+        "bbox": [
+            507,
+            772,
+            843,
+            788
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The Cross-lingual Alignment is a core challenge for cross-lingual transfer. Shi et al. (2022) proposed a series of prompts to elicit models to generate CoT answers in specific language Native-CoT, and in English En-CoT and Translate-CoT (more detailed in Table 1).",
+        "bbox": [
+            507,
+            791,
+            884,
+            887
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Later, Qin et al. (2023) proposed a method based on two phases: Cross-lingual alignment prompt and",
+        "bbox": [
+            507,
+            889,
+            882,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "1230",
+        "bbox": [
+            482,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Native-CoT in this example in Chinese",
+        "bbox": [
+            132,
+            82,
+            374,
+            93
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "问题：利亚有32块巧克力，她妹妹有42块。如果她们吃了35块，她们一共还剩下多少块？",
+        "bbox": [
+            132,
+            93,
+            467,
+            118
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "答案: 让我们一步步思考",
+        "bbox": [
+            134,
+            118,
+            295,
+            131
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "En-CoT",
+        "bbox": [
+            134,
+            134,
+            181,
+            143
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "问题：利亚有32块巧克力，她妹妹有42块。如果她们吃了35块，她们一共还剩下多少块？",
+        "bbox": [
+            132,
+            143,
+            467,
+            168
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Answer: Let's think step by step",
+        "bbox": [
+            134,
+            170,
+            378,
+            181
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Translated-CoT (is the Native translated in En)",
+        "bbox": [
+            134,
+            186,
+            433,
+            198
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Question: Leah has 32 chocolates and her sister has 42. If they ate 35 pieces, how many pieces do they have left?",
+        "bbox": [
+            132,
+            199,
+            468,
+            234
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Answer: Let's think step by step",
+        "bbox": [
+            132,
+            235,
+            378,
+            246
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Different types of input prompts in order to elicit Chain-of-Thought reasoning process. Specifically, given a problem in Chinese, the following prompts are Native-CoT and En-CoT, the original question in Chinese with elicitation in Chinese and English; for Translated-CoT, the question is in English and consequently a step-by-step solution in English.",
+        "bbox": [
+            112,
+            261,
+            487,
+            361
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "task-specific solver prompting. This approach uses two separate steps, as shown in Table 2, in order to handle input and output in different languages.",
+        "bbox": [
+            112,
+            386,
+            485,
+            434
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Cross-CoT First-Step",
+        "bbox": [
+            134,
+            445,
+            287,
+            457
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Please act as an expert in multi-lingual understanding in [Specific Language  $L_{s}$ ]. Question: [Given sentence  $X$  in  $L_{s}$ ] Let's understand the task in [Target Language  $L_{t}$ ] step-by-step!",
+        "bbox": [
+            132,
+            458,
+            468,
+            516
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Cross-CoT Second-Step",
+        "bbox": [
+            134,
+            519,
+            294,
+            530
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "After understanding, you should act as an expert in mathematics in [Language  $L_{t}$ ].",
+        "bbox": [
+            132,
+            531,
+            467,
+            552
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Let's resolve the task you understand above step-by-step!",
+        "bbox": [
+            132,
+            555,
+            467,
+            577
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Cross-lingual Prompt proposed in (Qin et al., 2023). By setting an input language and a target language, the prompt is divided into two phases: in phase one, there is the alignment of the different languages, and in phase two, there is the solving mechanism for the specific language.",
+        "bbox": [
+            112,
+            595,
+            487,
+            681
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Although this second approach overcomes the limitations of Shi et al. (2022)'s work, the two-step prompting could be more laborious and challenging, and there is no exchange of information during the multi-step reasoning process between the different chains as the final outputs are estimated using a voting heuristic.",
+        "bbox": [
+            112,
+            697,
+            487,
+            808
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.2 Self-consistent Parallel Chain-of-Thoughts",
+        "text_level": 1,
+        "bbox": [
+            112,
+            820,
+            484,
+            835
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In our work, we propose Cross-ToT, a prompting method that can handle different languages in a parallel way. Furthermore, through a mechanism inspired by Tree-of-Thoughts prompting techniques (Yao et al., 2023), our method elicits the LLM to",
+        "bbox": [
+            112,
+            841,
+            487,
+            920
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "deliver the generation of the answer in a sequence of intermediate steps that do not provide independent parallel answers but deliver collaborative Self-consistent reasoned steps until arriving at a final answer.",
+        "bbox": [
+            507,
+            84,
+            882,
+            162
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Our Proposal",
+        "text_level": 1,
+        "bbox": [
+            527,
+            173,
+            621,
+            185
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Simulate the collaboration of  $\\{n\\}$  mathematicians answering a question in their mother tongue:  $L_{1}, L_{2}, \\ldots$  and  $L_{n}$ . They all start Step1 from a separate thought process, step by step, each explaining their thought process. Following Step1, each expert refines and develops their thought process by comparing themselves with others. This process continues until a definitive answer to the question is obtained. Question: [Question in Language  $L_{1}$ ] Answer: [num].",
+        "bbox": [
+            526,
+            186,
+            865,
+            328
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Input-prompt for MSGM task. In Cross-ToT, we elicit the model to produce multi-step reasoning processes in different languages. We specifically prompt to start from separate reasoning and collaborate step-by-step. (We propose similar pattern for other tasks as described in Appendix A)",
+        "bbox": [
+            507,
+            340,
+            882,
+            426
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Our Cross-ToT shown in Table 3 elicits the LLM to generate different paths as shown in Figure 1, achieving significant improvements in accuracy as discussed in Section 4.",
+        "bbox": [
+            507,
+            444,
+            880,
+            506
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Experiments",
+        "text_level": 1,
+        "bbox": [
+            507,
+            520,
+            653,
+            536
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Data",
+        "text_level": 1,
+        "bbox": [
+            509,
+            546,
+            593,
+            558
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In order to observe the Cross-lingual abilities of LLMs, we used GSM8K (Cobbe et al., 2021), XNLI (Conneau et al., 2018), and PAWS-X (Yang et al., 2019), XCOPA (Ponti et al., 2020).",
+        "bbox": [
+            507,
+            565,
+            882,
+            630
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Understanding tasks In order to assess Crosslingual comprehension abilities, we used XNLI (Conneau et al., 2018) and PAWS-X. The first is an extension of Stanford Natural Language Inference (SNLI) (Bowman et al., 2015) across 15 languages and is based on one premise and one hypothesis and requires the model to determine whether the hypothesis is entailed, contradicted, or neutral conditioned on the premise in 15 different languages, and we utilize the accuracy score for evaluation. The second, Paraphrase Adversaries from Word Scrambling (PAWS-X) (Yang et al., 2019), contains two sentences and requires the model to judge whether they paraphrase each other in seven languages.",
+        "bbox": [
+            507,
+            639,
+            882,
+            879
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Commonsense Reasoning task The Cross-lingual Choice of Plausible Alternatives (XCOPA)",
+        "bbox": [
+            507,
+            889,
+            882,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "1231",
+        "bbox": [
+            482,
+            928,
+            517,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "table",
+        "img_path": "images/a5845f37440cadf04f9445120e84ade6a846ab528149afccb9cf7e6519f6ae22.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>de</td><td>zh</td><td>fr</td><td>ru</td><td>sw</td><td>es</td><td>bn</td><td>ja</td><td>te</td><td>th</td><td>Avg</td></tr><tr><td>GPT-3 (text-davinci-002)*</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>Direct (Shi et al., 2022)</td><td>14.8</td><td>18.0</td><td>16.8</td><td>12.4</td><td>8.8</td><td>17.2</td><td>4.4</td><td>11.2</td><td>0.8</td><td>8.8</td><td>11.3</td></tr><tr><td>Native-CoT (Shi et al., 2022)</td><td>36.0</td><td>40.0</td><td>37.6</td><td>28.4</td><td>11.2</td><td>40.4</td><td>6.4</td><td>26.0</td><td>0.4</td><td>10.8</td><td>23.7</td></tr><tr><td>En-CoT (Shi et al., 2022)</td><td>44.0</td><td>40.8</td><td>46.0</td><td>28.4</td><td>20.8</td><td>44.8</td><td>9.6</td><td>32.4</td><td>5.6</td><td>19.6</td><td>29.2</td></tr><tr><td>Translate-En (Shi et al., 2022)</td><td>46.4</td><td>47.2</td><td>46.4</td><td>48.8</td><td>37.6</td><td>51.6</td><td>41.2</td><td>44.8</td><td>42.8</td><td>41.2</td><td>44.8</td></tr><tr><td>GPT-3.5 (gpt-3.5-turbo)</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>Direct (Qin et al., 2023)</td><td>56.0</td><td>60.0</td><td>62.0</td><td>62.0</td><td>48.0</td><td>61.2</td><td>33.6</td><td>52.8</td><td>7.6</td><td>42.2</td><td>48.5</td></tr><tr><td>Native-CoT (Qin et al., 2023)</td><td>70.0</td><td>59.6</td><td>64.4</td><td>62.4</td><td>54.0</td><td>70.4</td><td>26.4</td><td>64.4</td><td>40.0</td><td>59.6</td><td>57.1</td></tr><tr><td>En-CoT (Qin et al., 2023)</td><td>73.6</td><td>63.2</td><td>70.0</td><td>65.6</td><td>55.2</td><td>69.6</td><td>50.0</td><td>60.4</td><td>22.0</td><td>48.0</td><td>57.7</td></tr><tr><td>Translate-En (Qin et al., 2023)</td><td>75.6</td><td>71.6</td><td>72.4</td><td>72.8</td><td>69.6</td><td>74.4</td><td>66.4</td><td>66.0</td><td>58.0</td><td>57.6</td><td>68.4</td></tr><tr><td>Cross-CoT (Qin et al., 2023)</td><td>86.8</td><td>77.2</td><td>82.0</td><td>87.6</td><td>76.0</td><td>84.8</td><td>75.2</td><td>77.2</td><td>52.0</td><td>68.0</td><td>76.6</td></tr><tr><td>Cross-ToT</td><td>87.6</td><td>83.5</td><td>84.3</td><td>86.5</td><td>75.4</td><td>86.2</td><td>79.0</td><td>80.2</td><td>68.5</td><td>75.5</td><td>80.6</td></tr></table>",
+        "bbox": [
+            114,
+            80,
+            957,
+            296
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Table 4: Accuracies (%) on MGSM using the \"Direct\" prompt, i.e., question and answer in the original language; the \"Native-CoT\" prompt, i.e., question and answer CoT in the original language; the \"En-CoT\" prompt specific language question and answer CoT in English, the \"Translate-En\" prompt where the specific input is translated into English and the answer accordingly is in English. Moreover, Cross-CoT, as proposed by Qin et al. (2023), questions in a specific language and answers in different languages. Finally, Cross-ToT is explained in Section 2.2. (Our results are derived from the average of three running performances as detailed in Section 3.2)",
+        "bbox": [
+            110,
+            305,
+            882,
+            391
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "(Ponti et al., 2020) is based on one premise and two choices. It asks the model to choose which one is the result or cause of the premise. It covers 11 languages from 11 diverse families.",
+        "bbox": [
+            112,
+            416,
+            487,
+            481
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Arithmetic Reasoning task To evaluate the problem-solving abilities in Cross-lingual scenarios, we used the extension proposed by Shi et al. (2022), i.e., Multilingual Grade School Math (MGSM). Initially, Cobbe et al. (2021) proposed a benchmark of mathematical problems in English in GSM8K. Each example has the following structure: a mathematical problem in natural language and a target answer in Arabic number. Shi et al. (2022), in their contribution, i.e., MGSM, selected the first 250 examples from the official list of examples in GSM8K and translated them manually into 11 different languages, maintaining the structure of the input and output.",
+        "bbox": [
+            112,
+            491,
+            489,
+            717
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Evaluated Languages In our experiments, we propose an analysis of available languages that differ depending on the resources, we provide all details in Appendix A. Furthermore, as an additional experiment, we test the introduction of English.",
+        "bbox": [
+            112,
+            726,
+            489,
+            808
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2 Experimental Setup",
+        "text_level": 1,
+        "bbox": [
+            112,
+            819,
+            317,
+            835
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In order to conduct our study on robust models and have a term of comparison with the work proposed in (Shi et al., 2022; Qin et al., 2023), we use GPT-3.5; however, in future developments, we plan to scale the method to different models. Then, we",
+        "bbox": [
+            112,
+            841,
+            489,
+            920
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "systematically defined the input prompt in Table 3 for MGSM and in Appendix A for XNLI, PAWS-X, and XCOPA. In each particular experimental set-up, we modify the appropriate languages with  $L_{1}, L_{2}, \\ldots$  for the German",
+        "bbox": [
+            507,
+            416,
+            884,
+            495
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Following Wei et al. (2022); Kojima et al. (2023), we evaluate performance using the accuracy score. In particular, we compute the string matching between the final answers (see Figure 1 where the final outputs have the form of Answer: [num]) and the target values. The top-p parameter is set to 1 in all processes. We select the Prompting temperature [0, 1].",
+        "bbox": [
+            507,
+            497,
+            885,
+            625
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4 Main Results",
+        "text_level": 1,
+        "bbox": [
+            507,
+            638,
+            658,
+            653
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Mechanisms for delivering multistep-reasoned answers across languages can be empowered via Cross-ToT that align languages' Chain-of-Thoughts (CoT). Our approach based on a Tree-of-Thoughts-inspired prompting mechanism (see Figure 1) outperforms state-of-the-art prompting techniques on Arithmetic Reasoning tasks as shown in Table 4, and in Language Understanding tasks as shown in Figure 3 and finally in Commonsense Reasoning tasks as shown in Table 5. In particular, Cross-ToT elicit LLMs to produce different reasoning pathways that share the \"Thoughts\" during the steps and, at the same time, promote Self-",
+        "bbox": [
+            507,
+            664,
+            885,
+            873
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>1</sup>Although we do not observe perceptible changes in the order of languages present in the input prompt, we set as a first the language-related subset of the benchmark.",
+        "bbox": [
+            507,
+            883,
+            882,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "1232",
+        "bbox": [
+            482,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "correction of mistaken paths. In fact, during the steps of the CoT, information is swapped between the paths. This interaction delivers Self-consistent paths. Furthermore, in the prompt, we exemplified that the different paths must arrive at a shared and, consequently, unique by sharing the \"thought process\" (see the prompt in Table 3).",
+        "bbox": [
+            112,
+            84,
+            489,
+            197
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/ee369d3a74e3f9f75bfd0b17a314a2a6041ac05bebfa53a633a853bee0b0ab42.jpg",
+        "image_caption": [
+            "Figure 2: Accuracies  $(\\%)$  on MGSM using \"Cross-ToT\", \"Cross-ToT + English\" and in binary version \"Cross-ToT (English + Target Language)\"."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            119,
+            211,
+            497,
+            370
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Our approach outperforms the methods proposed in (Shi et al., 2022) that are yet surpassed by the Cross-CoT proposed by Qin et al. (2023). However, although Cross-CoT outperforms previous approaches, it is necessary to clarify which path, if any, leads to the correct reasoning (Section 5.3), whether the introduction of English can increase performance (Section 5.1) and finally the trade-off between the number of languages (in our case path) and the final results (Section 5.2).",
+        "bbox": [
+            112,
+            444,
+            489,
+            605
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5 Analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            618,
+            223,
+            634
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In this section, we explore the contribution of English in the Cross-lingual prompt (in Section 5.1), then study the impact of different languages on the final results (Section 5.2) and the reasoning evolution (Section 5.3) and close with an in-depth analysis of performance in different tasks in Section 5.4.",
+        "bbox": [
+            112,
+            645,
+            489,
+            756
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5.1 The English Matter",
+        "text_level": 1,
+        "bbox": [
+            112,
+            770,
+            315,
+            785
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Earlier works (Wei et al., 2022; Liu et al., 2023) have been showing that LLMs are able to deliver multi-step reasoning answers on arithmetic tasks, focusing mainly on English. Therefore, we observe whether introducing English into our input-prompts could increase downstream performance. Hence, we performed the setting proposed in Section 3.2 From the results obtained in Figure 2",
+        "bbox": [
+            112,
+            791,
+            489,
+            920
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/4f55074069a4a2e286b4f6136c5e2fb88327720d4af485c6ed67167af5a4bfc7.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            598,
+            93,
+            794,
+            231
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/d79a77d7d9e654f96929625067c709d45da7751ab0a80f0d60f48c2a89a3c9b8.jpg",
+        "image_caption": [
+            "Figure 3: Accuracies  $(\\%)$  on Language Understanding benchmarks XNLI and PAWS-X introduced in Section 3.1"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            598,
+            249,
+            791,
+            385
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "(green bar), it is possible to observe that the input-prompts empowered with English outperform the input-prompts empowered without English. This result suggests that the presence of one robust path, in this case, the English path, may influence the others in the final reasoning process. Indeed, assuming that the production of the intermediate steps is self-consistent, i.e., the paths do not disagree with each other, the additional language seems to influence performance positively. From the current results, adding a further language improves the robustness of the models.",
+        "bbox": [
+            507,
+            475,
+            884,
+            667
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "However, whether the performance is due to the number of languages or English is unclear. To observe the impact of adding a specific language in Section 5.2, we propose to reduce the number of languages in the presence and absence of English.",
+        "bbox": [
+            507,
+            669,
+            882,
+            749
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5.2 The Impact of the Languages",
+        "text_level": 1,
+        "bbox": [
+            507,
+            760,
+            786,
+            777
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "English seems to lead Cross-lingual reasoning on arithmetic tasks, as shown in Section 5.1. Hence, to observe the impact of the number of languages and one specific, i.e., English, we propose two further analyses:",
+        "bbox": [
+            507,
+            782,
+            880,
+            862
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Cross-ToT in low-resources scenarios Integrating more languages into Cross-lingual prompting leads to better overall performance. As already",
+        "bbox": [
+            507,
+            873,
+            882,
+            921
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "1233",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/92cccbdf2b5c0a3574bdec000194a8f37c43b54b4792d0c34415b587efe6cb00.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>et</td><td>ht</td><td>id</td><td>it</td><td>qu</td><td>sw</td><td>ta</td><td>th</td><td>tr</td><td>vi</td><td>zh</td><td>Avg</td></tr><tr><td>GPT-3 (text-davinci-002)*</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>Direct (Shi et al., 2022)</td><td>73.8</td><td>55.6</td><td>88.8</td><td>95.4</td><td>51.2</td><td>56.0</td><td>54.6</td><td>70.2</td><td>88.6</td><td>80.4</td><td>91.4</td><td>73.3</td></tr><tr><td>En-CoT (Shi et al., 2022)</td><td>88.8</td><td>79.6</td><td>91.4</td><td>96.6</td><td>52.2</td><td>67.4</td><td>55.8</td><td>84.2</td><td>91.2</td><td>86.6</td><td>93.4</td><td>80.7</td></tr><tr><td>GPT-3.5 (gpt-3.5-turbo)</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>Direct (Qin et al., 2023)</td><td>90.6</td><td>72.0</td><td>90.4</td><td>95.2</td><td>54.6</td><td>82.0</td><td>59.0</td><td>77.6</td><td>91.0</td><td>83.6</td><td>90.4</td><td>80.6</td></tr><tr><td>Translate-En (Qin et al., 2023)</td><td>88.2</td><td>79.4</td><td>90.8</td><td>94.4</td><td>50.0</td><td>77.6</td><td>87.0</td><td>82.2</td><td>87.8</td><td>88.4</td><td>92.2</td><td>83.5</td></tr><tr><td>Cross-CoT (Qin et al., 2023)</td><td>96.8</td><td>90.6</td><td>95.2</td><td>95.8</td><td>85.8</td><td>92.8</td><td>83.2</td><td>93.2</td><td>96.8</td><td>94.2</td><td>95.8</td><td>92.7</td></tr><tr><td>Cross-ToT</td><td>97.6</td><td>92.5</td><td>90.3</td><td>96.8</td><td>83.3</td><td>93.6</td><td>80.2</td><td>94.1</td><td>96.4</td><td>95.3</td><td>97.4</td><td></td></tr><tr><td>HUMAN (Ponti et al., 2020)</td><td>98.2</td><td>96.4</td><td>100.0</td><td>97.0</td><td>94.8</td><td>99.0</td><td>98.6</td><td>98.2</td><td>96.4</td><td>98.4</td><td>96.6</td><td>97.6</td></tr></table>",
+        "bbox": [
+            114,
+            80,
+            905,
+            208
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 5: Accuracies (%) of XCOPA.",
+        "bbox": [
+            371,
+            217,
+            623,
+            231
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "observed in (Shi et al., 2022; Qin et al., 2023), increasing the number of languages improves downstream performance, as shown in Figure 4 (average performances using the same setting proposed in Section 3.2).",
+        "bbox": [
+            112,
+            256,
+            487,
+            336
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "As shown in (Malkin et al., 2022; Blevins and Zettlemoyer, 2022), the performances of the Large Language Models are highly correlated with the percentage of pre-training data in each language.",
+        "bbox": [
+            112,
+            338,
+            485,
+            403
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Following the approach proposed in (Qin et al., 2023) and considering language distribution in the widely used multilingual pre-training dataset, which in our case is CommonCrawl (Common Crawl, 2021), we integrated languages in descending and ascending order based on their respective proportions (detailed in Table 12).",
+        "bbox": [
+            112,
+            405,
+            487,
+            516
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Figure 4 shows that adding more languages in high-resource contexts improves performance. However, when incorporating languages with limited resources, performance decreases as the number of languages increases (see low-resource in Table 4). Finally, adding English (the dominant percentage in standard corpora) to the prompting significantly enhances performance (see \"+\" English\" lines in Table 4).",
+        "bbox": [
+            112,
+            519,
+            487,
+            662
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "These findings emphasize that the number of integrated languages only partially determines the effectiveness of language integration. The amount of pre-training data for each language, especially for high-resource languages, plays a crucial role. Balancing multiple languages and considering available resources and impact is essential.",
+        "bbox": [
+            112,
+            665,
+            487,
+            778
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Cross-ToT in binary scenarios Moreover, we evaluate similar scenarios in low-resource settings and reproduce the same experiments using only two languages. In particular, we used the same setting proposed in Section 3.2 by including only the target language and English in the prompt (example prompt in Appendix 8).",
+        "bbox": [
+            112,
+            791,
+            487,
+            903
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "From the results shown in Figure 2 (grey bar),",
+        "bbox": [
+            131,
+            904,
+            487,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "using the target English-language tuple does not change the performance of high-resource languages. On the contrary, low-resource languages achieve significantly lower performance. This second finding reinforces what was said earlier about the experiments on prompt compositions.",
+        "bbox": [
+            507,
+            258,
+            884,
+            354
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/190bb14aa541118614d945eab32c3e3d6eb95a8938eadeb33f2e5e29b4f78a08.jpg",
+        "image_caption": [
+            "Figure 4: The impact of integrating languages in our Cross-ToT on the final performance. Following Table 12, we integrate languages from low-resources to high-resources and vice versa. We also propose the same experiments with the addition of English."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            549,
+            376,
+            892,
+            542
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.3 Reasoning Evolution",
+        "text_level": 1,
+        "bbox": [
+            507,
+            658,
+            719,
+            673
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We use the framework ROSCOE (Golovneva et al., 2023) to investigate why our approach works. Hence, we evaluate the quality of the reasoning paths (implementation described in Appendix B). As shown in Figure 5, our approach delivers reasoning with higher faithfulness, exhibiting better consistency with key steps during the reasoning process. Specifically, the faithfulness score increased by 4.5 points, indicating that the model better understood the problem statement and ensured a transparent inference chain without generating irrelevant or misused information. Furthermore, we observe improvements in the Informativeness metrics for \"Step\" and \"Chain\". It suggests that the models' reasoning, behind the alignment, could",
+        "bbox": [
+            505,
+            678,
+            884,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "1234",
+        "bbox": [
+            482,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "provide more well-grounded inference steps.",
+        "bbox": [
+            114,
+            85,
+            445,
+            99
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/e05f086b6c6abb6af40c2223c84ef713cb63a73ac8dcd700fded89754fd6c4f6.jpg",
+        "image_caption": [
+            "Figure 5: The analysis of reasoning quality between GPT-3.5 (Native-CoT) and CLP in (Qin et al., 2023) and our Cross-ToT"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            136,
+            122,
+            478,
+            272
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "XCOPA, XNLI and PAWS-X",
+        "text_level": 1,
+        "bbox": [
+            124,
+            367,
+            332,
+            381
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Simulate the collaboration of  $n$  person answering a question in their mother tongue:  $L_{1}$  and English. They all start Step1 from a separate thought process, step by step, each explaining their thought process. Following Step1, each expert refines and develops their thought process by comparing themselves with others. This process continues until a definitive answer to the question is obtained.",
+        "bbox": [
+            122,
+            384,
+            463,
+            575
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Basic Prompt",
+        "bbox": [
+            126,
+            577,
+            238,
+            590
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 6: Our prompting approach for XCOPA, XNLI and PAWS-X. List of the Basic Prompt is in Table 11",
+        "bbox": [
+            112,
+            604,
+            485,
+            633
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.4 The Cross-Reasoning in other tasks",
+        "text_level": 1,
+        "bbox": [
+            112,
+            663,
+            438,
+            678
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Furthermore, to scale our approach, we test the applicability of Cross-ToT on two different task types using the same structure adapted to them as in Table 7.",
+        "bbox": [
+            112,
+            684,
+            485,
+            747
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Understanding task We proposed our approach, Cross-ToT, on other multilingual reasoning datasets belonging to the understandings genre. As introduced in Section 3.2, we used XNLI (Conneau et al., 2018) and PAWS-X (Yang et al., 2019). As Figure 3 shows, Cross-ToT is able to perform better in most languages. Compared to the previous SOTA obtained in CLP (Qin et al., 2023). Thus, we observed average improvements of 3.2 points on XNLI and 2.5 points on PAWS-X.",
+        "bbox": [
+            112,
+            760,
+            489,
+            920
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Commonsense Reasoning task We have used our approach, Cross-ToT, to an additional dataset of multilingual commonsense reasoning, as introduced in Section 3.1. We used XCOPA as our benchmark. For comparison purposes, we considered CLP and Native-CoT proposed by Qin et al. (2023). In Figure 5, we can observe that our approach has outperformed previous methods in many languages.",
+        "bbox": [
+            507,
+            84,
+            884,
+            228
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "The results show the effective functionality of our Cross-ToT on different tasks. Although the method has shown appreciable increases, we continue the studies in Section 5.5 by observing whether adding in-context examples in the input-prompt can benefit LLMs.",
+        "bbox": [
+            507,
+            229,
+            884,
+            325
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.5 Other approaches",
+        "text_level": 1,
+        "bbox": [
+            507,
+            335,
+            697,
+            350
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Cross-ToT can be further empowered with in-context learning. In fact, as shown in Table 9, in-context learning (ICL) techniques have achieved performant results on the downstream performance of LLMs. In particular, in further exploration of Cross-ToT within ICL, we conducted different experiments.",
+        "bbox": [
+            507,
+            356,
+            882,
+            467
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "From Zero-to Few-shot In the first experiment, we sampled 50 random instances from MGSM. Then, we replicated the experiments proposed in Section 3.2. However, we constructed the prompt by merging instances in one-shot and three-shot settings. Table 9 shows that providing context makes the models more robust.",
+        "bbox": [
+            507,
+            476,
+            882,
+            587
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Performances Other Models Cross-ToT does not outperform other approaches in open-source models with fewer parameters. Table 10 shows the performances of Llama-2-13B (Touvron et al., 2023) and Bloomz-7B (Muennighoff et al., 2022). We hypothesize that these performances are due to the misleading behaviors observed in (Wei et al., 2023) prompting CoT in models with less than 100 billion parameters. In future developments, we will continue to investigate this phenomenon.",
+        "bbox": [
+            507,
+            596,
+            882,
+            757
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "6 Related Work",
+        "text_level": 1,
+        "bbox": [
+            507,
+            768,
+            665,
+            783
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Large Language Models (LLMs) with billions of parameters demonstrate in-context learning and few-shot learning abilities (Brown et al., 2020; Wei et al., 2022; Min et al., 2022) to guide LLMs to generate desired task responses, marking the advent of the prompting era and surpassing the age of the intermediate steps in algorithmic and structured reasoning (Roy and Roth, 2015; Ling et al.,",
+        "bbox": [
+            507,
+            791,
+            882,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "1235",
+        "bbox": [
+            482,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "2017). Nevertheless, early works challenged the efficacy of few-shot techniques for empowering the prompting phase and downstream performances. In particular, Yao et al. (2023) refined the original idea of Chain-of-Thought (CoT) (Wei et al., 2022) by considering various reasoning paths as well known as Tree-of-Thought.",
+        "bbox": [
+            112,
+            84,
+            487,
+            197
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "The traditional and derivated CoT mechanisms have achieved considerable success but are limited to generating answers within a single language (i.e., English). Shi et al. (2022) proposed a multilingual evaluation that Qin et al. (2023) extended to cross-lingual scenarios. In particular, Qin et al. (2023) proposed a prompt mechanism to handle requests in any language and generate CoT specifically in English. This approach, which in our construct we called Cross-CoT has been proposed both single-phase, i.e., as a single prompt (CLP) also adopted by (Huang et al., 2023) and multiphase (CLPS) i.e., characterized by self-consistent prompts that follow the prompting methodology proposed in (Qiao et al., 2023). Although the mechanism achieves state-of-the-art cross-linguistic reasoning steps, the single-phase promoting underperforms in low-resources languages and the multiphase prompting characterized by a series of cascading prompts is supported far away from the zero-shot chain-of-thought concept.",
+        "bbox": [
+            115,
+            199,
+            489,
+            535
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In our work, we propose a method of CoT reasoning inspired. Specifically, we elicit the crosslingual generation of a series of parallel Crosslingual reasoning paths using a single prompt. In fact, our method is inspired by the Tree-of-Thoughts approach proposed by (Yao et al., 2023). Hence, in a different way from previous approaches, our technique generates shared parallel reasoning paths that share the \"thoughts process\" delivering Self-consistent answers and reducing reasoning steps. Our work goes beyond in the following ways:",
+        "bbox": [
+            112,
+            538,
+            489,
+            731
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Proposal of novel zero-shot prompting methods in cross-lingual scenarios characterized by low-resource and high-resource languages.",
+            "- Studying cross-lingual multi-step reasoning mechanisms using arithmetic reasoning tasks.",
+            "- In-depth study of the reasoning pathways provided by our prompting approach (impact of the number of languages and strongly high-resource languages)."
+        ],
+        "bbox": [
+            136,
+            745,
+            487,
+            920
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "- Experiments on effective functioning in commonsense reasoning and language understanding tasks.",
+        "bbox": [
+            531,
+            84,
+            882,
+            131
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "7 Future Works",
+        "text_level": 1,
+        "bbox": [
+            507,
+            143,
+            665,
+            159
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In future work, we intend to incorporate smaller-scale Language Models (SLMs) into our evaluations. However, the ability to produce multi-step reasoned answers is limited in SLMs. To address this, a range of techniques are emerging to align and transfer reasoning abilities between LLMs and SLMs (Ranaldi and Freitas, 2024).",
+        "bbox": [
+            507,
+            170,
+            882,
+            281
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Our aim is to enhance current alignment pipelines (Ranaldi et al., 2023; Ranaldi and Pucci, 2023a) to enable cross-lingual reasoning capabilities across different languages and scenarios. Including methods that emphasize the importance of language structure (Zanzotto et al., 2020) and uphold the foundational pillars of the NLP ecosystem (Ranaldi and Pucci, 2023b).",
+        "bbox": [
+            507,
+            282,
+            882,
+            411
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "8 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            423,
+            640,
+            437
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Chain-of-Thought is an outstanding prompting technique. However, the imbalance of languages in pre-training data does not always produce robust results. Different state-of-the-art works have proposed cross-lingual techniques to align performances obtained in different languages. They are limited to handling one language at a time or proposing multiple prompting stages, making them difficult to manage. In this paper, we propose Cross-ToT, a prompting technique to elicit multi-step reasoning abilities in Cross-lingual scenarios. Hence, we elicit models to deliver answers in a Self-consistent way, collaborating to the final answer. We have shown the functionality of our Cross-ToT through performance improvements obtained in a multilingual mathematical problem task. In addition, we have demonstrated the scalability in tasks related to commonsense reasoning and language understanding. Finally, we conducted a series of in-depth analyses in which we measured the impact brought about by low-resource vs. high-resource languages and the inclusion of English. Our contribution aims to propose more robust models that can break down issues arising from language barriers and provide more reliable results.",
+        "bbox": [
+            507,
+            449,
+            882,
+            851
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            507,
+            863,
+            613,
+            878
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Due to the limitations imposed by the evaluation benchmarks and the cost of the OpenAI API, we",
+        "bbox": [
+            507,
+            889,
+            880,
+            920
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "1236",
+        "bbox": [
+            482,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "conducted tests on 16 languages in total, which only scratches the surface of the world's vast array of languages. Furthermore, our approach is based on English. It should be evaluated whether the model written in the language of the task can lead to better performance and how best to construct instructions in each language. Furthermore, we only tested the effectiveness of our method on GPT-based models (gpt-3.5-turbo). In the future, it will be worthwhile to study the generality of our model on more models, such as PaLM and Llama-2-70.",
+        "bbox": [
+            112,
+            84,
+            489,
+            261
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Ethics Statemets",
+        "text_level": 1,
+        "bbox": [
+            114,
+            275,
+            263,
+            288
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In our work, ethical topics were not addressed. The data used comes from open-source benchmarks, and statistics on language differences in commonly used pre-training data were obtained from official sources without touching on issues related to gender, sex, or race differences.",
+        "bbox": [
+            112,
+            302,
+            490,
+            398
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            426,
+            213,
+            441
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Terra Blevins and Luke Zettlemoyer. 2022. Language contamination helps explain the cross-lingual capabilities of english pretrained models.",
+            "Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. 2015. A large annotated corpus for learning natural language inference. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 632-642, Lisbon, Portugal. Association for Computational Linguistics.",
+            "Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners.",
+            "Sebastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, Harsha Nori, Hamid Palangi, Marco Tulio Ribeiro, and Yi Zhang. 2023. Sparks of artificial general intelligence: Early experiments with gpt-4.",
+            "Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek"
+        ],
+        "bbox": [
+            115,
+            450,
+            489,
+            920
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan First, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. 2022. Palm: Scaling language modeling with pathways.",
+            "Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. Training verifiers to solve math word problems.",
+            "Common crawl. 2021. Common crawl 2021. Web. Accessed: 2023-12-12.",
+            "Alexis Conneau, Rudy Rinott, Guillaume Lample, Adina Williams, Samuel Bowman, Holger Schwenk, and Veselin Stoyanov. 2018. XNLI: Evaluating crosslingual sentence representations. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2475-2485, Brussels, Belgium. Association for Computational Linguistics.",
+            "Vedant Gaur and Nikunj Saunshi. 2023. Reasoning in large language models through symbolic math word problems. In *Findings of the Association for Computational Linguistics: ACL* 2023, pages 5889-5903, Toronto, Canada. Association for Computational Linguistics.",
+            "Olga Golovneva, Moya Chen, Spencer Poff, Martin Corredor, Luke Zettlemoyer, Maryam Fazel-Zarandi, and Asli Celikyilmaz. 2023. Roscoe: A suite of metrics for scoring step-by-step reasoning.",
+            "Haoyang Huang, Tianyi Tang, Dongdong Zhang, Wayne Xin Zhao, Ting Song, Yan Xia, and Furu Wei. 2023. Not all languages are created equal in llms: Improving multilingual capability by cross-lingualthought prompting.",
+            "Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2023. Large language models are zero-shot reasoners.",
+            "Wang Ling, Dani Yogatama, Chris Dyer, and Phil Blun-. som. 2017. Program induction by rationale generation: Learning to solve and explain algebraic word problems. In Proceedings of the 55th Annual Meet- ing of the Association for Computational Linguistics"
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            920
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "1237",
+        "bbox": [
+            482,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "(Volume 1: Long Papers), pages 158-167, Vancouver, Canada. Association for Computational Linguistics.",
+            "Hanmeng Liu, Ruoxi Ning, Zhiyang Teng, Jian Liu, Qiji Zhou, and Yue Zhang. 2023. Evaluating the logical reasoning ability of chatgpt and gpt-4.",
+            "Dan Malkin, Tomasz Limisiewicz, and Gabriel Stanovsky. 2022. A balanced data approach for evaluating cross-lingual transfer: Mapping the linguistic blood bank. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4903–4915, Seattle, United States. Association for Computational Linguistics.",
+            "Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettlemoyer. 2022. Rethinking the role of demonstrations: What makes in-context learning work? In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 11048-11064, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.",
+            "Niklas Muennighoff, Thomas Wang, Lintang Sutawika, Adam Roberts, Stella Biderman, Teven Le Scao, M Saiful Bari, Sheng Shen, Zheng-Xin Yong, Hailey Schoelkopf, et al. 2022. Crosslingual generalization through multitask finetuning. arXiv preprint arXiv:2211.01786.",
+            "OpenAI. 2023. Gpt-4 technical report.",
+            "Edoardo Maria Ponti, Goran Glavas, Olga Majewska, Qianchu Liu, Ivan Vulic, and Anna Korhonen. 2020. XCOPA: A multilingual dataset for causal commonsense reasoning. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2362-2376, Online. Association for Computational Linguistics.",
+            "Shuofei Qiao, Yixin Ou, Ningyu Zhang, Xiang Chen, Yunzhi Yao, Shumin Deng, Chuanqi Tan, Fei Huang, and Huajun Chen. 2023. Reasoning with language model prompting: A survey. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5368-5393, Toronto, Canada. Association for Computational Linguistics.",
+            "Libo Qin, Qiguang Chen, Fuxuan Wei, Shijue Huang, and Wanxiang Che. 2023. Cross-lingual prompting: Improving zero-shot chain-of-thought reasoning across languages.",
+            "Leonardo Ranaldi and Andre Freitas. 2024. Aligning large and small language models via chain-of-thought reasoning. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1812-1827, St. Julian's, Malta. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            115,
+            85,
+            487,
+            919
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Leonardo Ranaldi and Giulia Pucci. 2023a. Does the English matter? elicit cross-lingual abilities of large language models. In Proceedings of the 3rd Workshop on Multi-lingual Representation Learning (MRL), pages 173–183, Singapore. Association for Computational Linguistics.",
+            "Leonardo Ranaldi and Giulia Pucci. 2023b. Knowing knowledge: Epistemological study of knowledge in transformers. Applied Sciences, 13(2).",
+            "Leonardo Ranaldi, Giulia Pucci, and Andre Freitas. 2023. Empowering cross-lingual abilities of instruction-tuned large language models by translation-following demonstrations.",
+            "Subhro Roy and Dan Roth. 2015. Solving general arithmetic word problems. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 1743-1752, Lisbon, Portugal. Association for Computational Linguistics.",
+            "Freda Shi, Mirac Suzgun, Markus Freitag, Xuezhi Wang, Suraj Srivats, Soroush Vosoughi, Hyung Won Chung, Yi Tay, Sebastian Ruder, Denny Zhou, Dipanjan Das, and Jason Wei. 2022. Language models are multilingual chain-of-thought reasoners.",
+            "Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. 2023. Llama 2: Open foundation and finetuned chat models.",
+            "Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, Ed H. Chi, Tatsunori Hashimoto, Oriol Vinyals, Percy Liang, Jeff Dean, and William Fedus. 2022. Emergent abilities of large language models.",
+            "Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. 2023. Chain-of-thought prompting elicits reasoning in large language models."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            919
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "1238",
+        "bbox": [
+            482,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Yinfei Yang, Yuan Zhang, Chris Tar, and Jason Baldridge. 2019. PAWS-X: A cross-lingual adversarial dataset for paraphrase identification. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3687-3692, Hong Kong, China. Association for Computational Linguistics.",
+            "Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L. Griffiths, Yuan Cao, and Karthik Narasimhan. 2023. Tree of thoughts: Deliberate problem solving with large language models.",
+            "Fabio Massimo Zanzotto, Andrea Santilli, Leonardo Ranaldi, Dario Onorati, Pierfrancesco Tommasino, and Francesca Fallucchi. 2020. KERMIT: Complementing transformer architectures with encoders of explicit syntactic interpretations. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 256-267, Online. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            381
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "1239",
+        "bbox": [
+            482,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A Prompt",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            220,
+            99
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we analyze our prompting approach, i.e., Cross-ToT, in different tasks. In Figure 1 we have shown the input-prompt for the MGSM (Cobbe et al., 2021). Here, we show the prompt framework for the other tasks:",
+        "bbox": [
+            112,
+            108,
+            487,
+            187
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "XCOPA, XNLI and PAWS-X",
+        "text_level": 1,
+        "bbox": [
+            124,
+            196,
+            332,
+            210
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Simulate the collaboration of  $n$  person answering a question in their mother tongue:  $L_{1}$  and English. They all start Step1 from a separate thought process, step by step, each explaining their thought process. Following Step1, each expert refines and develops their thought process by comparing themselves with others. This process continues until a definitive answer to the question is obtained.",
+        "bbox": [
+            122,
+            212,
+            463,
+            403
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Basic Prompt",
+        "text_level": 1,
+        "bbox": [
+            124,
+            406,
+            240,
+            420
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Table 7: Our prompting approach for XCOPA, XNLI and PAWS-X. List of the Basic Prompt is in Table 11",
+        "bbox": [
+            112,
+            432,
+            485,
+            462
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Furthermore, in Section 5.1, we proposed an experiment based on a prompt with only two languages as follows:",
+        "bbox": [
+            112,
+            479,
+            487,
+            527
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Binary Cross-ToT",
+        "text_level": 1,
+        "bbox": [
+            132,
+            536,
+            257,
+            548
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Simulate the collaboration of 2 mathematicians answering a question in their mother tongue:  $L_{1}$  and English. They all start Step1 from a separate thought process, step by step, each explaining their thought process. Following Step1, each expert refines and develops their thought process by comparing themselves with others. This process continues until a definitive answer to the question is obtained. Question: [Question in Language  $L_{1}$ ] Answer: [num].",
+        "bbox": [
+            131,
+            549,
+            472,
+            690
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Table 8: Our prompting approach for experiment proposed in Section 5.1 regarding MGSM and binary trees",
+        "bbox": [
+            112,
+            703,
+            487,
+            733
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B Reasoning Chain",
+        "text_level": 1,
+        "bbox": [
+            114,
+            757,
+            300,
+            772
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B.1 Chain-of-Thought Quality Scoring Implementation",
+        "text_level": 1,
+        "bbox": [
+            114,
+            781,
+            435,
+            812
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "The ROSCOE framework (Golovneva et al., 2023) incorporates multiple chain-of-thought quality metrics, with the reasoning alignment vector  $\\alpha$  that is",
+        "bbox": [
+            112,
+            818,
+            487,
+            881
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nr _ {\\text {a l i g n}} (h \\rightarrow s) = \\left\\{\\alpha_ {1}, \\alpha_ {2}, \\dots , \\alpha_ {N} \\right\\} \\in [ 0, 1 ] ^ {N} \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            129,
+            885,
+            487,
+            920
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "from the  $N$ -step hypothesis  $h = \\{h_i\\}_{i=1}^N$  to the source input  $s$  of length  $T$ , where  $\\alpha_i$  are defined as:  $r_{align}(h_i \\to s) = \\frac{1 + \\max_{j=1}^{T} \\cos(h_i, s_j)}{2}$",
+        "bbox": [
+            507,
+            83,
+            882,
+            140
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Faithfulness score The Faithfulness  $(F)$  score is calculated based on the alignment between the hypothesis steps  $h$  and the source sentences  $s$ . It represents the average reasoning alignment score over the steps of reasoning:",
+        "bbox": [
+            507,
+            145,
+            882,
+            227
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nF = \\frac {1}{N} \\sum_ {i = 1} ^ {N} r _ {\\text {a l i g n}} \\left(h _ {i} \\rightarrow s\\right) \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            591,
+            237,
+            882,
+            280
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "The Faithfulness score serves as a measure to assess whether the model misconstrued the problem in the statement or if the reasoning chain is characterized by ambiguity, unimportance, or the misuse of information.",
+        "bbox": [
+            507,
+            291,
+            882,
+            370
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Informativeness Informativeness-Step (Info-Step) measures the utilization of facts from the original text  $s$  in the reasoning steps  $h$ :",
+        "bbox": [
+            507,
+            380,
+            882,
+            429
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nI n f o _ {S t e p} = \\frac {1}{2 T} \\sum_ {t = 1} ^ {T} r _ {\\text {a l i g n}} \\left(s _ {t} \\rightarrow h\\right) + \\frac {1}{2} F \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            526,
+            439,
+            882,
+            482
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Info-Step assigns a higher score to reasoning steps that strongly align with the source, showing the capacity to which the generated hypothesis includes the information from the source. Conversely, a lower Info-Step score means reasoning steps unrelated to the source sentences or overlooking the provided information in the context.",
+        "bbox": [
+            507,
+            493,
+            882,
+            605
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Informativeness Chain Like the Info-Step metric, the InformativenessChain (Info-Chain) metric estimates the degree of concordance between the hypothesis chain and the source. The calculation is as follows:",
+        "bbox": [
+            507,
+            614,
+            882,
+            693
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nI n f o _ {C h a i n} = \\frac {1 + \\cos (h , s)}{2} \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            588,
+            703,
+            882,
+            733
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Missing Step The Missing Step (Miss-Step) metric is introduced to estimate any significant lacking steps, which examines the alignment between the reference reasoning text  $r = \\{r_i\\}^K$  and the hypothesis  $h$ . A miss-step is needed to meticulously assess each step in the reference and verify the existence of a similar step in the hypothesis. The metric is computed as:",
+        "bbox": [
+            507,
+            744,
+            882,
+            871
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\text {M i s s - S t e p} = \\min  _ {i = 1} ^ {K} (\\mathrm {r} - \\operatorname {a l i g n} (r _ {i}, h)). \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            569,
+            884,
+            882,
+            915
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "1240",
+        "bbox": [
+            482,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "C Other Results",
+        "text_level": 1,
+        "bbox": [
+            114,
+            173,
+            273,
+            187
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "table",
+        "img_path": "images/54462bf8b9370401ee5b6bec61f8f1177b366f1f71efc2ff5a6b54ad08401367.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td># of shot-Cross-ToT</td><td>de</td><td>zh</td><td>fr</td><td>ru</td><td>sw</td><td>es</td><td>bn</td><td>ja</td><td>te</td><td>th</td><td>Avg</td></tr><tr><td>0-shot</td><td>86.5</td><td>84.2</td><td>83.9</td><td>83.2</td><td>74.3</td><td>84.4</td><td>78.7</td><td>79.8</td><td>68.7</td><td>74.6</td><td>79.8</td></tr><tr><td>1-shot</td><td>87.2</td><td>84.9</td><td>85.8</td><td>85.3</td><td>76.4</td><td>85.2</td><td>81.2</td><td>81.3</td><td>70.5</td><td>75.5</td><td>79.9</td></tr><tr><td>3-shot</td><td>88.4</td><td>85.7</td><td>87.2</td><td>87.5</td><td>77.3</td><td>87.3</td><td>82.3</td><td>81.5</td><td>70.3</td><td>76.9</td><td>83.4</td></tr></table>",
+        "bbox": [
+            119,
+            193,
+            877,
+            263
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "table",
+        "img_path": "images/a8b0259cf9d9539a74a2250169d66179897fdf56537029ca9bcbdf7cf4688a7e.jpg",
+        "table_caption": [
+            "Table 9: Accuracies (%) on MGSM using zero-shot, one-shot and three-shot"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>et</td><td>ht</td><td>id</td><td>it</td><td>qu</td><td>sw</td><td>ta</td><td>th</td><td>tr</td><td>vi</td><td>zh</td><td>Avg</td></tr><tr><td colspan=\"13\">Bloomz-7B (Muennighoff et al., 2022)</td></tr><tr><td>En-CoT</td><td>21.8</td><td>24.2</td><td>50.6</td><td>41.6</td><td>41.4</td><td>48.6</td><td>53.8</td><td>38.4</td><td>37.6</td><td>47.0</td><td>64.2</td><td>42.7</td></tr><tr><td>CLP (Qin et al., 2023)</td><td>49.0</td><td>49.6</td><td>58.0</td><td>48.8</td><td>50.6</td><td>47.6</td><td>57.8</td><td>52.0</td><td>50.2</td><td>45.2</td><td>54.2</td><td>51.2</td></tr><tr><td>Cross-ToT</td><td>48.0</td><td>47.3</td><td>58.2</td><td>47.8</td><td>49.3</td><td>46.4</td><td>55.2</td><td>53.1</td><td>50.8</td><td>44.2</td><td>50.3</td><td>49.5</td></tr><tr><td colspan=\"13\">llama-2-13B (Touvron et al., 2023)</td></tr><tr><td>En-CoT</td><td>39.6</td><td>32.5</td><td>58.4</td><td>55.8</td><td>47.2</td><td>34.6</td><td>47.4</td><td>33.2</td><td>43.0</td><td>59.6</td><td>50.4</td><td>45.6</td></tr><tr><td>CLP (Qin et al., 2023)</td><td>44.8</td><td>48.2</td><td>64.4</td><td>70.2</td><td>46.6</td><td>47.0</td><td>47.8</td><td>46.4</td><td>51.2</td><td>58.8</td><td>51.4</td><td>52.4</td></tr><tr><td>Cross-ToT</td><td>43.3</td><td>49.1</td><td>61.5</td><td>65.8</td><td>44.4</td><td>46.6</td><td>43.7</td><td>42.2</td><td>49.5</td><td>55.2</td><td>48.2</td><td>50.6</td></tr></table>",
+        "bbox": [
+            115,
+            312,
+            932,
+            426
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "D Prompt Table",
+        "text_level": 1,
+        "bbox": [
+            114,
+            458,
+            272,
+            475
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "table",
+        "img_path": "images/bc7cea3dc59fcdd7f8eedbc7612e8dfe5612b63a0ac070982f061db71f1bed4e.jpg",
+        "table_caption": [
+            "Table 10: Comparison of smaller open-source models on XCOPA."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Benchmark</td><td>#Test</td><td>Basic Prompt</td></tr><tr><td>MGSM</td><td>250</td><td>Question: {problem}</td></tr><tr><td>XCOPA</td><td>200</td><td>Here is a premise: {premise}. What is the {question}? Help me pick the more plausible option: -choice1: {choice1}, -choice2: {choice2}</td></tr><tr><td>XNLI</td><td>200</td><td>{premise}. Based on the previous passage, is it true that {hypothesis}? Yes, No, or Maybe?</td></tr><tr><td>PAWS-X</td><td>200</td><td>Sentence 1: {sentence1} Sentence 2: {sentence2} Question: Does Sentence 1 paraphrase Sentence 2? Yes or No?</td></tr></table>",
+        "bbox": [
+            127,
+            481,
+            867,
+            573
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "E Number of Languages",
+        "text_level": 1,
+        "bbox": [
+            114,
+            657,
+            344,
+            674
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "table",
+        "img_path": "images/cb20a16a60f372aea9fa0d48d7940abc0a638eeefd8ac1e7508290af184fbca6.jpg",
+        "table_caption": [
+            "Table 11: The basic prompt of each benchmark. #Test denotes the number of instances in the test set that we randomly selected due to the cost constraint excepted for MGSM."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Language</td><td>Percentage</td></tr><tr><td>English (en)</td><td>46.3%</td></tr><tr><td>Russian (ru)</td><td>6.0%</td></tr><tr><td>German (de)</td><td>5.4%</td></tr><tr><td>Chinese (zh)</td><td>5.3%</td></tr><tr><td>French (fr)</td><td>4.4%</td></tr><tr><td>Japanese (ja)</td><td>4.3%</td></tr><tr><td>Spanish (es)</td><td>4.2%</td></tr><tr><td>Other</td><td>23.1%</td></tr></table>",
+        "bbox": [
+            401,
+            678,
+            596,
+            793
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Table 12: Language distribution of CommonCrawl (Common Crawl, 2021).",
+        "bbox": [
+            238,
+            803,
+            752,
+            818
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "1241",
+        "bbox": [
+            482,
+            928,
+            517,
+            940
+        ],
+        "page_idx": 12
+    }
+]

2024/A Tree-of-Thoughts to Broaden Multi-step Reasoning across Languages/cf769ade-bf6e-4081-af4e-c7800f171438_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e8b692e1046f3dc32cb7f0f3cab09be6280a38537208ab22b9db35d076d15bf4
+size 987756

2024/A Tree-of-Thoughts to Broaden Multi-step Reasoning across Languages/full.md

@@ -0,0 +1,388 @@
+# A Tree-of-Thoughts to Broaden Multi-step Reasoning across Languages
+
+Leonardo Ranaldi $(\diamond)$ , Giulia Pucci  $(\star)$ , Federico Ranaldi  $(\diamond)$ , Elena Sofia Ruzzetti  $(\diamond)$ , Fabio Massimo Zanzotto $(\diamond)$
+
+$(\diamond)$  Human-Centric ART Group,
+
+Dep. of Enterprise Engineering, University of Rome Tor Vergata, Italy
+
+$(\star)$ Department of Computing Science, University of Aberdeen, UK
+
+[first_name].[last_name]@uniroma2.it
+
+# Abstract
+
+Reasoning methods, best exemplified by the well-known Chain-of-Thought (CoT), empower the reasoning abilities of Large Language Models (LLMs) by eliciting them to solve complex tasks in a step-by-step manner. Although they are achieving significant success, the ability to deliver multi-step reasoning remains limited to English because of the imbalance in the distribution of pre-training data, which makes other languages a barrier.
+
+In this paper, we propose Cross-lingual Tree-of-Thoughts (Cross-ToT), a method for aligning Cross-lingual CoT reasoning across languages. The proposed method, through a self-consistent cross-lingual prompting mechanism inspired by the Tree-of-Thoughts approach, provides multi-step reasoning paths in different languages that, during the steps, lead to the final solution. Experimental evaluations show that our method significantly outperforms existing prompting methods by reducing the number of interactions and achieving state-of-the-art performance.
+
+# 1 Introduction
+
+Chain-of-Thought (CoT) prompting elicits Large Language Models (LLMs) to break down a reasoning task towards a sequence of intermediate steps (Wei et al., 2022). Previous works have demonstrated that LLMs achieve impressive performances in zero-shot learning scenarios without the need to modify the model parameters during the training and testing process. In particular, by appending to the prompt "Let's think step by step!" (Kojima et al., 2023) LLMs with at least several billions of parameters, such as GPTs family (OpenAI, 2023) or PaLM (Chowdhery et al., 2022), deliver multi-step controlled reasoning, achieving promising results across commonsense (Bubeck et al., 2023), symbolic and mathematical reasoning datasets (Gaur and Saunshi, 2023; Liu et al., 2023).
+
+Although the performances seem promising, they are only firmly established in English. This
+
+poses a barrier to generalizing current CoT techniques to different languages. Hence, despite the remarkable success of zero-shot CoT techniques, the reasoning abilities of LLMs still struggle to generalize to different languages. Shi et al. (2022) introduced the first multilingual benchmark to assess LLMs' mathematical reasoning abilities using prompts in different languages. Qin et al. (2023) propose task-specific solver prompting, using a succession of prompts, elicit the LLMs to understand questions and deliver CoT answers in different languages. However, these strategies require two-step prompts, which goes against the zero-shot approach.
+
+In this paper, we propose Cross-lingual Tree-of-Thoughts (Cross-ToT), a method for aligning Cross-lingual CoT reasoning across languages by proposing a Cross-lingual Alignment prompt to elicit the model to deliver a Self-consistent Chain-of-Thought. Our method is inspired by the Tree-of-Thoughts (ToT) prompting (Yao et al., 2023) that asks LLMs to perform decision-making by considering multiple different reasoning paths (CoTs). In particular, our Cross-ToT is a ToT-style prompting to deliver the reasoning process in different languages that, step-by-step, converge to a single final solution. The inherent insight is that as the different paths of thought evolve, the relationships between the different languages are inherently grasped via Self-consistent Chains-of-Thought. This leads to the target research questions, which are the focus of this paper:
+
+RQ1: Are LLMs able to deliver Cross-lingual multi-step reasoned answers?
+
+$RQ2$ : Are the different paths of ToT evolving Self-correcting each other?
+
+$RQ3$ : What is the role of English in Cross-lingual scenarios?
+
+To answer these questions, we propose Cross-ToT, a novel Cross-lingual prompting strategy that aims to bridge the gap across different
+
+![](images/ea195486126cd5fbe63d99bb4827d5b88b7fdb2929b011908c4c890d1eefd1ce.jpg)  
+Figure 1: Our Cross-ToT elicits the LLM to generate step-by-step Cross-lingual reasoning. Furthermore, different pathways are developed during these reasoning steps. This mechanism develops the Chains-of-Thoughts in a Self-consistent way, streaming with the different pathways.
+
+languages. In particular, using the prompt shown in Figure 1, we elicit the model to deliver different CoT reasoning steps in different languages that converge to the final solution step-by-step. We test our method on GPT-3.5 and conduct an extensive analysis using Multilingual Grade School Math (MGSM) (Shi et al., 2022), Cross-lingual Natural Language Inference (XNLI) (Conneau et al., 2018), and Cross-lingual Paraphrase Adversaries Scrambling (PAWS-X) (Yang et al., 2019), Cross-lingual Choice of Plausible Alternatives (XCOPA) (Ponti et al., 2020) across different languages. Experimental results reveal that our method, based on a single prompt, outperforms the baselines and achieves the SOTA performance on different languages in different tasks. The main contributions of this work are concluded as follows:
+
+- We introduce Cross-ToT, which is a novel Cross-lingual prompting mechanism that stimulates the model to produce parallel CoT reasoning processes across different languages;  
+- We show that our Cross-ToT is Self-consistent and allows the integration of reasoning paths between different languages;  
+- Extensive evaluations on different languages
+
+demonstrate that our Cross-ToT can effectively improve the performance of crosslingual CoTs and achieve SOTA performance.
+
+- Finally, we show that introducing English in our prompting technique plays a beneficial role in improving downstream performance.
+
+# 2 Cross-lingual Multi-step Reasoning
+
+To elicit the multi-step reasoning abilities of LLMs in Cross-lingual scenarios, we propose Cross-ToT, which is a Cross-lingual Alignment Chain-of-Thought as a solution. In particular, our method overcomes the Multi-lingual and Cross-lingual approaches introduced in Section 2.1. In fact, our approach elicits the LLMs to deliver Self-consistent Parallel Chain-of-Thoughts, introduced in Section 2.2.
+
+# 2.1 Chain-of-Thought Across Languages
+
+The Cross-lingual Alignment is a core challenge for cross-lingual transfer. Shi et al. (2022) proposed a series of prompts to elicit models to generate CoT answers in specific language Native-CoT, and in English En-CoT and Translate-CoT (more detailed in Table 1).
+
+Later, Qin et al. (2023) proposed a method based on two phases: Cross-lingual alignment prompt and
+
+Native-CoT in this example in Chinese
+
+问题：利亚有32块巧克力，她妹妹有42块。如果她们吃了35块，她们一共还剩下多少块？
+
+答案: 让我们一步步思考
+
+En-CoT
+
+问题：利亚有32块巧克力，她妹妹有42块。如果她们吃了35块，她们一共还剩下多少块？
+
+Answer: Let's think step by step
+
+Translated-CoT (is the Native translated in En)
+
+Question: Leah has 32 chocolates and her sister has 42. If they ate 35 pieces, how many pieces do they have left?
+
+Answer: Let's think step by step
+
+Table 1: Different types of input prompts in order to elicit Chain-of-Thought reasoning process. Specifically, given a problem in Chinese, the following prompts are Native-CoT and En-CoT, the original question in Chinese with elicitation in Chinese and English; for Translated-CoT, the question is in English and consequently a step-by-step solution in English.
+
+task-specific solver prompting. This approach uses two separate steps, as shown in Table 2, in order to handle input and output in different languages.
+
+Cross-CoT First-Step
+
+Please act as an expert in multi-lingual understanding in [Specific Language  $L_{s}$ ]. Question: [Given sentence  $X$  in  $L_{s}$ ] Let's understand the task in [Target Language  $L_{t}$ ] step-by-step!
+
+Cross-CoT Second-Step
+
+After understanding, you should act as an expert in mathematics in [Language  $L_{t}$ ].
+
+Let's resolve the task you understand above step-by-step!
+
+Table 2: Cross-lingual Prompt proposed in (Qin et al., 2023). By setting an input language and a target language, the prompt is divided into two phases: in phase one, there is the alignment of the different languages, and in phase two, there is the solving mechanism for the specific language.
+
+Although this second approach overcomes the limitations of Shi et al. (2022)'s work, the two-step prompting could be more laborious and challenging, and there is no exchange of information during the multi-step reasoning process between the different chains as the final outputs are estimated using a voting heuristic.
+
+# 2.2 Self-consistent Parallel Chain-of-Thoughts
+
+In our work, we propose Cross-ToT, a prompting method that can handle different languages in a parallel way. Furthermore, through a mechanism inspired by Tree-of-Thoughts prompting techniques (Yao et al., 2023), our method elicits the LLM to
+
+deliver the generation of the answer in a sequence of intermediate steps that do not provide independent parallel answers but deliver collaborative Self-consistent reasoned steps until arriving at a final answer.
+
+# Our Proposal
+
+Simulate the collaboration of  $\{n\}$  mathematicians answering a question in their mother tongue:  $L_{1}, L_{2}, \ldots$  and  $L_{n}$ . They all start Step1 from a separate thought process, step by step, each explaining their thought process. Following Step1, each expert refines and develops their thought process by comparing themselves with others. This process continues until a definitive answer to the question is obtained. Question: [Question in Language  $L_{1}$ ] Answer: [num].
+
+Table 3: Input-prompt for MSGM task. In Cross-ToT, we elicit the model to produce multi-step reasoning processes in different languages. We specifically prompt to start from separate reasoning and collaborate step-by-step. (We propose similar pattern for other tasks as described in Appendix A)
+
+Our Cross-ToT shown in Table 3 elicits the LLM to generate different paths as shown in Figure 1, achieving significant improvements in accuracy as discussed in Section 4.
+
+# 3 Experiments
+
+# 3.1 Data
+
+In order to observe the Cross-lingual abilities of LLMs, we used GSM8K (Cobbe et al., 2021), XNLI (Conneau et al., 2018), and PAWS-X (Yang et al., 2019), XCOPA (Ponti et al., 2020).
+
+Understanding tasks In order to assess Crosslingual comprehension abilities, we used XNLI (Conneau et al., 2018) and PAWS-X. The first is an extension of Stanford Natural Language Inference (SNLI) (Bowman et al., 2015) across 15 languages and is based on one premise and one hypothesis and requires the model to determine whether the hypothesis is entailed, contradicted, or neutral conditioned on the premise in 15 different languages, and we utilize the accuracy score for evaluation. The second, Paraphrase Adversaries from Word Scrambling (PAWS-X) (Yang et al., 2019), contains two sentences and requires the model to judge whether they paraphrase each other in seven languages.
+
+Commonsense Reasoning task The Cross-lingual Choice of Plausible Alternatives (XCOPA)
+
+<table><tr><td>Model</td><td>de</td><td>zh</td><td>fr</td><td>ru</td><td>sw</td><td>es</td><td>bn</td><td>ja</td><td>te</td><td>th</td><td>Avg</td></tr><tr><td>GPT-3 (text-davinci-002)*</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>Direct (Shi et al., 2022)</td><td>14.8</td><td>18.0</td><td>16.8</td><td>12.4</td><td>8.8</td><td>17.2</td><td>4.4</td><td>11.2</td><td>0.8</td><td>8.8</td><td>11.3</td></tr><tr><td>Native-CoT (Shi et al., 2022)</td><td>36.0</td><td>40.0</td><td>37.6</td><td>28.4</td><td>11.2</td><td>40.4</td><td>6.4</td><td>26.0</td><td>0.4</td><td>10.8</td><td>23.7</td></tr><tr><td>En-CoT (Shi et al., 2022)</td><td>44.0</td><td>40.8</td><td>46.0</td><td>28.4</td><td>20.8</td><td>44.8</td><td>9.6</td><td>32.4</td><td>5.6</td><td>19.6</td><td>29.2</td></tr><tr><td>Translate-En (Shi et al., 2022)</td><td>46.4</td><td>47.2</td><td>46.4</td><td>48.8</td><td>37.6</td><td>51.6</td><td>41.2</td><td>44.8</td><td>42.8</td><td>41.2</td><td>44.8</td></tr><tr><td>GPT-3.5 (gpt-3.5-turbo)</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>Direct (Qin et al., 2023)</td><td>56.0</td><td>60.0</td><td>62.0</td><td>62.0</td><td>48.0</td><td>61.2</td><td>33.6</td><td>52.8</td><td>7.6</td><td>42.2</td><td>48.5</td></tr><tr><td>Native-CoT (Qin et al., 2023)</td><td>70.0</td><td>59.6</td><td>64.4</td><td>62.4</td><td>54.0</td><td>70.4</td><td>26.4</td><td>64.4</td><td>40.0</td><td>59.6</td><td>57.1</td></tr><tr><td>En-CoT (Qin et al., 2023)</td><td>73.6</td><td>63.2</td><td>70.0</td><td>65.6</td><td>55.2</td><td>69.6</td><td>50.0</td><td>60.4</td><td>22.0</td><td>48.0</td><td>57.7</td></tr><tr><td>Translate-En (Qin et al., 2023)</td><td>75.6</td><td>71.6</td><td>72.4</td><td>72.8</td><td>69.6</td><td>74.4</td><td>66.4</td><td>66.0</td><td>58.0</td><td>57.6</td><td>68.4</td></tr><tr><td>Cross-CoT (Qin et al., 2023)</td><td>86.8</td><td>77.2</td><td>82.0</td><td>87.6</td><td>76.0</td><td>84.8</td><td>75.2</td><td>77.2</td><td>52.0</td><td>68.0</td><td>76.6</td></tr><tr><td>Cross-ToT</td><td>87.6</td><td>83.5</td><td>84.3</td><td>86.5</td><td>75.4</td><td>86.2</td><td>79.0</td><td>80.2</td><td>68.5</td><td>75.5</td><td>80.6</td></tr></table>
+
+Table 4: Accuracies (%) on MGSM using the "Direct" prompt, i.e., question and answer in the original language; the "Native-CoT" prompt, i.e., question and answer CoT in the original language; the "En-CoT" prompt specific language question and answer CoT in English, the "Translate-En" prompt where the specific input is translated into English and the answer accordingly is in English. Moreover, Cross-CoT, as proposed by Qin et al. (2023), questions in a specific language and answers in different languages. Finally, Cross-ToT is explained in Section 2.2. (Our results are derived from the average of three running performances as detailed in Section 3.2)
+
+(Ponti et al., 2020) is based on one premise and two choices. It asks the model to choose which one is the result or cause of the premise. It covers 11 languages from 11 diverse families.
+
+Arithmetic Reasoning task To evaluate the problem-solving abilities in Cross-lingual scenarios, we used the extension proposed by Shi et al. (2022), i.e., Multilingual Grade School Math (MGSM). Initially, Cobbe et al. (2021) proposed a benchmark of mathematical problems in English in GSM8K. Each example has the following structure: a mathematical problem in natural language and a target answer in Arabic number. Shi et al. (2022), in their contribution, i.e., MGSM, selected the first 250 examples from the official list of examples in GSM8K and translated them manually into 11 different languages, maintaining the structure of the input and output.
+
+Evaluated Languages In our experiments, we propose an analysis of available languages that differ depending on the resources, we provide all details in Appendix A. Furthermore, as an additional experiment, we test the introduction of English.
+
+# 3.2 Experimental Setup
+
+In order to conduct our study on robust models and have a term of comparison with the work proposed in (Shi et al., 2022; Qin et al., 2023), we use GPT-3.5; however, in future developments, we plan to scale the method to different models. Then, we
+
+systematically defined the input prompt in Table 3 for MGSM and in Appendix A for XNLI, PAWS-X, and XCOPA. In each particular experimental set-up, we modify the appropriate languages with  $L_{1}, L_{2}, \ldots$  for the German
+
+Following Wei et al. (2022); Kojima et al. (2023), we evaluate performance using the accuracy score. In particular, we compute the string matching between the final answers (see Figure 1 where the final outputs have the form of Answer: [num]) and the target values. The top-p parameter is set to 1 in all processes. We select the Prompting temperature [0, 1].
+
+# 4 Main Results
+
+Mechanisms for delivering multistep-reasoned answers across languages can be empowered via Cross-ToT that align languages' Chain-of-Thoughts (CoT). Our approach based on a Tree-of-Thoughts-inspired prompting mechanism (see Figure 1) outperforms state-of-the-art prompting techniques on Arithmetic Reasoning tasks as shown in Table 4, and in Language Understanding tasks as shown in Figure 3 and finally in Commonsense Reasoning tasks as shown in Table 5. In particular, Cross-ToT elicit LLMs to produce different reasoning pathways that share the "Thoughts" during the steps and, at the same time, promote Self-
+
+correction of mistaken paths. In fact, during the steps of the CoT, information is swapped between the paths. This interaction delivers Self-consistent paths. Furthermore, in the prompt, we exemplified that the different paths must arrive at a shared and, consequently, unique by sharing the "thought process" (see the prompt in Table 3).
+
+![](images/ee369d3a74e3f9f75bfd0b17a314a2a6041ac05bebfa53a633a853bee0b0ab42.jpg)  
+Figure 2: Accuracies  $(\%)$  on MGSM using "Cross-ToT", "Cross-ToT + English" and in binary version "Cross-ToT (English + Target Language)".
+
+Our approach outperforms the methods proposed in (Shi et al., 2022) that are yet surpassed by the Cross-CoT proposed by Qin et al. (2023). However, although Cross-CoT outperforms previous approaches, it is necessary to clarify which path, if any, leads to the correct reasoning (Section 5.3), whether the introduction of English can increase performance (Section 5.1) and finally the trade-off between the number of languages (in our case path) and the final results (Section 5.2).
+
+# 5 Analysis
+
+In this section, we explore the contribution of English in the Cross-lingual prompt (in Section 5.1), then study the impact of different languages on the final results (Section 5.2) and the reasoning evolution (Section 5.3) and close with an in-depth analysis of performance in different tasks in Section 5.4.
+
+# 5.1 The English Matter
+
+Earlier works (Wei et al., 2022; Liu et al., 2023) have been showing that LLMs are able to deliver multi-step reasoning answers on arithmetic tasks, focusing mainly on English. Therefore, we observe whether introducing English into our input-prompts could increase downstream performance. Hence, we performed the setting proposed in Section 3.2 From the results obtained in Figure 2
+
+![](images/4f55074069a4a2e286b4f6136c5e2fb88327720d4af485c6ed67167af5a4bfc7.jpg)
+
+![](images/d79a77d7d9e654f96929625067c709d45da7751ab0a80f0d60f48c2a89a3c9b8.jpg)  
+Figure 3: Accuracies  $(\%)$  on Language Understanding benchmarks XNLI and PAWS-X introduced in Section 3.1
+
+(green bar), it is possible to observe that the input-prompts empowered with English outperform the input-prompts empowered without English. This result suggests that the presence of one robust path, in this case, the English path, may influence the others in the final reasoning process. Indeed, assuming that the production of the intermediate steps is self-consistent, i.e., the paths do not disagree with each other, the additional language seems to influence performance positively. From the current results, adding a further language improves the robustness of the models.
+
+However, whether the performance is due to the number of languages or English is unclear. To observe the impact of adding a specific language in Section 5.2, we propose to reduce the number of languages in the presence and absence of English.
+
+# 5.2 The Impact of the Languages
+
+English seems to lead Cross-lingual reasoning on arithmetic tasks, as shown in Section 5.1. Hence, to observe the impact of the number of languages and one specific, i.e., English, we propose two further analyses:
+
+Cross-ToT in low-resources scenarios Integrating more languages into Cross-lingual prompting leads to better overall performance. As already
+
+<table><tr><td>Model</td><td>et</td><td>ht</td><td>id</td><td>it</td><td>qu</td><td>sw</td><td>ta</td><td>th</td><td>tr</td><td>vi</td><td>zh</td><td>Avg</td></tr><tr><td>GPT-3 (text-davinci-002)*</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>Direct (Shi et al., 2022)</td><td>73.8</td><td>55.6</td><td>88.8</td><td>95.4</td><td>51.2</td><td>56.0</td><td>54.6</td><td>70.2</td><td>88.6</td><td>80.4</td><td>91.4</td><td>73.3</td></tr><tr><td>En-CoT (Shi et al., 2022)</td><td>88.8</td><td>79.6</td><td>91.4</td><td>96.6</td><td>52.2</td><td>67.4</td><td>55.8</td><td>84.2</td><td>91.2</td><td>86.6</td><td>93.4</td><td>80.7</td></tr><tr><td>GPT-3.5 (gpt-3.5-turbo)</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>Direct (Qin et al., 2023)</td><td>90.6</td><td>72.0</td><td>90.4</td><td>95.2</td><td>54.6</td><td>82.0</td><td>59.0</td><td>77.6</td><td>91.0</td><td>83.6</td><td>90.4</td><td>80.6</td></tr><tr><td>Translate-En (Qin et al., 2023)</td><td>88.2</td><td>79.4</td><td>90.8</td><td>94.4</td><td>50.0</td><td>77.6</td><td>87.0</td><td>82.2</td><td>87.8</td><td>88.4</td><td>92.2</td><td>83.5</td></tr><tr><td>Cross-CoT (Qin et al., 2023)</td><td>96.8</td><td>90.6</td><td>95.2</td><td>95.8</td><td>85.8</td><td>92.8</td><td>83.2</td><td>93.2</td><td>96.8</td><td>94.2</td><td>95.8</td><td>92.7</td></tr><tr><td>Cross-ToT</td><td>97.6</td><td>92.5</td><td>90.3</td><td>96.8</td><td>83.3</td><td>93.6</td><td>80.2</td><td>94.1</td><td>96.4</td><td>95.3</td><td>97.4</td><td></td></tr><tr><td>HUMAN (Ponti et al., 2020)</td><td>98.2</td><td>96.4</td><td>100.0</td><td>97.0</td><td>94.8</td><td>99.0</td><td>98.6</td><td>98.2</td><td>96.4</td><td>98.4</td><td>96.6</td><td>97.6</td></tr></table>
+
+Table 5: Accuracies (%) of XCOPA.
+
+observed in (Shi et al., 2022; Qin et al., 2023), increasing the number of languages improves downstream performance, as shown in Figure 4 (average performances using the same setting proposed in Section 3.2).
+
+As shown in (Malkin et al., 2022; Blevins and Zettlemoyer, 2022), the performances of the Large Language Models are highly correlated with the percentage of pre-training data in each language.
+
+Following the approach proposed in (Qin et al., 2023) and considering language distribution in the widely used multilingual pre-training dataset, which in our case is CommonCrawl (Common Crawl, 2021), we integrated languages in descending and ascending order based on their respective proportions (detailed in Table 12).
+
+Figure 4 shows that adding more languages in high-resource contexts improves performance. However, when incorporating languages with limited resources, performance decreases as the number of languages increases (see low-resource in Table 4). Finally, adding English (the dominant percentage in standard corpora) to the prompting significantly enhances performance (see "+" English" lines in Table 4).
+
+These findings emphasize that the number of integrated languages only partially determines the effectiveness of language integration. The amount of pre-training data for each language, especially for high-resource languages, plays a crucial role. Balancing multiple languages and considering available resources and impact is essential.
+
+Cross-ToT in binary scenarios Moreover, we evaluate similar scenarios in low-resource settings and reproduce the same experiments using only two languages. In particular, we used the same setting proposed in Section 3.2 by including only the target language and English in the prompt (example prompt in Appendix 8).
+
+From the results shown in Figure 2 (grey bar),
+
+using the target English-language tuple does not change the performance of high-resource languages. On the contrary, low-resource languages achieve significantly lower performance. This second finding reinforces what was said earlier about the experiments on prompt compositions.
+
+![](images/190bb14aa541118614d945eab32c3e3d6eb95a8938eadeb33f2e5e29b4f78a08.jpg)  
+Figure 4: The impact of integrating languages in our Cross-ToT on the final performance. Following Table 12, we integrate languages from low-resources to high-resources and vice versa. We also propose the same experiments with the addition of English.
+
+# 5.3 Reasoning Evolution
+
+We use the framework ROSCOE (Golovneva et al., 2023) to investigate why our approach works. Hence, we evaluate the quality of the reasoning paths (implementation described in Appendix B). As shown in Figure 5, our approach delivers reasoning with higher faithfulness, exhibiting better consistency with key steps during the reasoning process. Specifically, the faithfulness score increased by 4.5 points, indicating that the model better understood the problem statement and ensured a transparent inference chain without generating irrelevant or misused information. Furthermore, we observe improvements in the Informativeness metrics for "Step" and "Chain". It suggests that the models' reasoning, behind the alignment, could
+
+provide more well-grounded inference steps.
+
+![](images/e05f086b6c6abb6af40c2223c84ef713cb63a73ac8dcd700fded89754fd6c4f6.jpg)  
+Figure 5: The analysis of reasoning quality between GPT-3.5 (Native-CoT) and CLP in (Qin et al., 2023) and our Cross-ToT
+
+# XCOPA, XNLI and PAWS-X
+
+Simulate the collaboration of  $n$  person answering a question in their mother tongue:  $L_{1}$  and English. They all start Step1 from a separate thought process, step by step, each explaining their thought process. Following Step1, each expert refines and develops their thought process by comparing themselves with others. This process continues until a definitive answer to the question is obtained.
+
+Basic Prompt
+
+Table 6: Our prompting approach for XCOPA, XNLI and PAWS-X. List of the Basic Prompt is in Table 11
+
+# 5.4 The Cross-Reasoning in other tasks
+
+Furthermore, to scale our approach, we test the applicability of Cross-ToT on two different task types using the same structure adapted to them as in Table 7.
+
+Understanding task We proposed our approach, Cross-ToT, on other multilingual reasoning datasets belonging to the understandings genre. As introduced in Section 3.2, we used XNLI (Conneau et al., 2018) and PAWS-X (Yang et al., 2019). As Figure 3 shows, Cross-ToT is able to perform better in most languages. Compared to the previous SOTA obtained in CLP (Qin et al., 2023). Thus, we observed average improvements of 3.2 points on XNLI and 2.5 points on PAWS-X.
+
+Commonsense Reasoning task We have used our approach, Cross-ToT, to an additional dataset of multilingual commonsense reasoning, as introduced in Section 3.1. We used XCOPA as our benchmark. For comparison purposes, we considered CLP and Native-CoT proposed by Qin et al. (2023). In Figure 5, we can observe that our approach has outperformed previous methods in many languages.
+
+The results show the effective functionality of our Cross-ToT on different tasks. Although the method has shown appreciable increases, we continue the studies in Section 5.5 by observing whether adding in-context examples in the input-prompt can benefit LLMs.
+
+# 5.5 Other approaches
+
+Cross-ToT can be further empowered with in-context learning. In fact, as shown in Table 9, in-context learning (ICL) techniques have achieved performant results on the downstream performance of LLMs. In particular, in further exploration of Cross-ToT within ICL, we conducted different experiments.
+
+From Zero-to Few-shot In the first experiment, we sampled 50 random instances from MGSM. Then, we replicated the experiments proposed in Section 3.2. However, we constructed the prompt by merging instances in one-shot and three-shot settings. Table 9 shows that providing context makes the models more robust.
+
+Performances Other Models Cross-ToT does not outperform other approaches in open-source models with fewer parameters. Table 10 shows the performances of Llama-2-13B (Touvron et al., 2023) and Bloomz-7B (Muennighoff et al., 2022). We hypothesize that these performances are due to the misleading behaviors observed in (Wei et al., 2023) prompting CoT in models with less than 100 billion parameters. In future developments, we will continue to investigate this phenomenon.
+
+# 6 Related Work
+
+Large Language Models (LLMs) with billions of parameters demonstrate in-context learning and few-shot learning abilities (Brown et al., 2020; Wei et al., 2022; Min et al., 2022) to guide LLMs to generate desired task responses, marking the advent of the prompting era and surpassing the age of the intermediate steps in algorithmic and structured reasoning (Roy and Roth, 2015; Ling et al.,
+
+2017). Nevertheless, early works challenged the efficacy of few-shot techniques for empowering the prompting phase and downstream performances. In particular, Yao et al. (2023) refined the original idea of Chain-of-Thought (CoT) (Wei et al., 2022) by considering various reasoning paths as well known as Tree-of-Thought.
+
+The traditional and derivated CoT mechanisms have achieved considerable success but are limited to generating answers within a single language (i.e., English). Shi et al. (2022) proposed a multilingual evaluation that Qin et al. (2023) extended to cross-lingual scenarios. In particular, Qin et al. (2023) proposed a prompt mechanism to handle requests in any language and generate CoT specifically in English. This approach, which in our construct we called Cross-CoT has been proposed both single-phase, i.e., as a single prompt (CLP) also adopted by (Huang et al., 2023) and multiphase (CLPS) i.e., characterized by self-consistent prompts that follow the prompting methodology proposed in (Qiao et al., 2023). Although the mechanism achieves state-of-the-art cross-linguistic reasoning steps, the single-phase promoting underperforms in low-resources languages and the multiphase prompting characterized by a series of cascading prompts is supported far away from the zero-shot chain-of-thought concept.
+
+In our work, we propose a method of CoT reasoning inspired. Specifically, we elicit the crosslingual generation of a series of parallel Crosslingual reasoning paths using a single prompt. In fact, our method is inspired by the Tree-of-Thoughts approach proposed by (Yao et al., 2023). Hence, in a different way from previous approaches, our technique generates shared parallel reasoning paths that share the "thoughts process" delivering Self-consistent answers and reducing reasoning steps. Our work goes beyond in the following ways:
+
+- Proposal of novel zero-shot prompting methods in cross-lingual scenarios characterized by low-resource and high-resource languages.  
+- Studying cross-lingual multi-step reasoning mechanisms using arithmetic reasoning tasks.  
+- In-depth study of the reasoning pathways provided by our prompting approach (impact of the number of languages and strongly high-resource languages).
+
+- Experiments on effective functioning in commonsense reasoning and language understanding tasks.
+
+# 7 Future Works
+
+In future work, we intend to incorporate smaller-scale Language Models (SLMs) into our evaluations. However, the ability to produce multi-step reasoned answers is limited in SLMs. To address this, a range of techniques are emerging to align and transfer reasoning abilities between LLMs and SLMs (Ranaldi and Freitas, 2024).
+
+Our aim is to enhance current alignment pipelines (Ranaldi et al., 2023; Ranaldi and Pucci, 2023a) to enable cross-lingual reasoning capabilities across different languages and scenarios. Including methods that emphasize the importance of language structure (Zanzotto et al., 2020) and uphold the foundational pillars of the NLP ecosystem (Ranaldi and Pucci, 2023b).
+
+# 8 Conclusion
+
+Chain-of-Thought is an outstanding prompting technique. However, the imbalance of languages in pre-training data does not always produce robust results. Different state-of-the-art works have proposed cross-lingual techniques to align performances obtained in different languages. They are limited to handling one language at a time or proposing multiple prompting stages, making them difficult to manage. In this paper, we propose Cross-ToT, a prompting technique to elicit multi-step reasoning abilities in Cross-lingual scenarios. Hence, we elicit models to deliver answers in a Self-consistent way, collaborating to the final answer. We have shown the functionality of our Cross-ToT through performance improvements obtained in a multilingual mathematical problem task. In addition, we have demonstrated the scalability in tasks related to commonsense reasoning and language understanding. Finally, we conducted a series of in-depth analyses in which we measured the impact brought about by low-resource vs. high-resource languages and the inclusion of English. Our contribution aims to propose more robust models that can break down issues arising from language barriers and provide more reliable results.
+
+# Limitations
+
+Due to the limitations imposed by the evaluation benchmarks and the cost of the OpenAI API, we
+
+conducted tests on 16 languages in total, which only scratches the surface of the world's vast array of languages. Furthermore, our approach is based on English. It should be evaluated whether the model written in the language of the task can lead to better performance and how best to construct instructions in each language. Furthermore, we only tested the effectiveness of our method on GPT-based models (gpt-3.5-turbo). In the future, it will be worthwhile to study the generality of our model on more models, such as PaLM and Llama-2-70.
+
+# Ethics Statemets
+
+In our work, ethical topics were not addressed. The data used comes from open-source benchmarks, and statistics on language differences in commonly used pre-training data were obtained from official sources without touching on issues related to gender, sex, or race differences.
+
+# References
+
+Terra Blevins and Luke Zettlemoyer. 2022. Language contamination helps explain the cross-lingual capabilities of english pretrained models.  
+Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. 2015. A large annotated corpus for learning natural language inference. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 632-642, Lisbon, Portugal. Association for Computational Linguistics.  
+Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners.  
+Sebastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, Harsha Nori, Hamid Palangi, Marco Tulio Ribeiro, and Yi Zhang. 2023. Sparks of artificial general intelligence: Early experiments with gpt-4.  
+Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek
+
+Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan First, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. 2022. Palm: Scaling language modeling with pathways.  
+Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. Training verifiers to solve math word problems.  
+Common crawl. 2021. Common crawl 2021. Web. Accessed: 2023-12-12.  
+Alexis Conneau, Rudy Rinott, Guillaume Lample, Adina Williams, Samuel Bowman, Holger Schwenk, and Veselin Stoyanov. 2018. XNLI: Evaluating crosslingual sentence representations. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2475-2485, Brussels, Belgium. Association for Computational Linguistics.  
+Vedant Gaur and Nikunj Saunshi. 2023. Reasoning in large language models through symbolic math word problems. In *Findings of the Association for Computational Linguistics: ACL* 2023, pages 5889-5903, Toronto, Canada. Association for Computational Linguistics.  
+Olga Golovneva, Moya Chen, Spencer Poff, Martin Corredor, Luke Zettlemoyer, Maryam Fazel-Zarandi, and Asli Celikyilmaz. 2023. Roscoe: A suite of metrics for scoring step-by-step reasoning.  
+Haoyang Huang, Tianyi Tang, Dongdong Zhang, Wayne Xin Zhao, Ting Song, Yan Xia, and Furu Wei. 2023. Not all languages are created equal in llms: Improving multilingual capability by cross-lingualthought prompting.  
+Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2023. Large language models are zero-shot reasoners.  
+Wang Ling, Dani Yogatama, Chris Dyer, and Phil Blun-. som. 2017. Program induction by rationale generation: Learning to solve and explain algebraic word problems. In Proceedings of the 55th Annual Meet- ing of the Association for Computational Linguistics
+
+(Volume 1: Long Papers), pages 158-167, Vancouver, Canada. Association for Computational Linguistics.  
+Hanmeng Liu, Ruoxi Ning, Zhiyang Teng, Jian Liu, Qiji Zhou, and Yue Zhang. 2023. Evaluating the logical reasoning ability of chatgpt and gpt-4.  
+Dan Malkin, Tomasz Limisiewicz, and Gabriel Stanovsky. 2022. A balanced data approach for evaluating cross-lingual transfer: Mapping the linguistic blood bank. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4903–4915, Seattle, United States. Association for Computational Linguistics.  
+Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettlemoyer. 2022. Rethinking the role of demonstrations: What makes in-context learning work? In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 11048-11064, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Niklas Muennighoff, Thomas Wang, Lintang Sutawika, Adam Roberts, Stella Biderman, Teven Le Scao, M Saiful Bari, Sheng Shen, Zheng-Xin Yong, Hailey Schoelkopf, et al. 2022. Crosslingual generalization through multitask finetuning. arXiv preprint arXiv:2211.01786.  
+OpenAI. 2023. Gpt-4 technical report.  
+Edoardo Maria Ponti, Goran Glavas, Olga Majewska, Qianchu Liu, Ivan Vulic, and Anna Korhonen. 2020. XCOPA: A multilingual dataset for causal commonsense reasoning. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2362-2376, Online. Association for Computational Linguistics.  
+Shuofei Qiao, Yixin Ou, Ningyu Zhang, Xiang Chen, Yunzhi Yao, Shumin Deng, Chuanqi Tan, Fei Huang, and Huajun Chen. 2023. Reasoning with language model prompting: A survey. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5368-5393, Toronto, Canada. Association for Computational Linguistics.  
+Libo Qin, Qiguang Chen, Fuxuan Wei, Shijue Huang, and Wanxiang Che. 2023. Cross-lingual prompting: Improving zero-shot chain-of-thought reasoning across languages.  
+Leonardo Ranaldi and Andre Freitas. 2024. Aligning large and small language models via chain-of-thought reasoning. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1812-1827, St. Julian's, Malta. Association for Computational Linguistics.
+
+Leonardo Ranaldi and Giulia Pucci. 2023a. Does the English matter? elicit cross-lingual abilities of large language models. In Proceedings of the 3rd Workshop on Multi-lingual Representation Learning (MRL), pages 173–183, Singapore. Association for Computational Linguistics.  
+Leonardo Ranaldi and Giulia Pucci. 2023b. Knowing knowledge: Epistemological study of knowledge in transformers. Applied Sciences, 13(2).  
+Leonardo Ranaldi, Giulia Pucci, and Andre Freitas. 2023. Empowering cross-lingual abilities of instruction-tuned large language models by translation-following demonstrations.  
+Subhro Roy and Dan Roth. 2015. Solving general arithmetic word problems. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 1743-1752, Lisbon, Portugal. Association for Computational Linguistics.  
+Freda Shi, Mirac Suzgun, Markus Freitag, Xuezhi Wang, Suraj Srivats, Soroush Vosoughi, Hyung Won Chung, Yi Tay, Sebastian Ruder, Denny Zhou, Dipanjan Das, and Jason Wei. 2022. Language models are multilingual chain-of-thought reasoners.  
+Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. 2023. Llama 2: Open foundation and finetuned chat models.  
+Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, Ed H. Chi, Tatsunori Hashimoto, Oriol Vinyals, Percy Liang, Jeff Dean, and William Fedus. 2022. Emergent abilities of large language models.  
+Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. 2023. Chain-of-thought prompting elicits reasoning in large language models.
+
+Yinfei Yang, Yuan Zhang, Chris Tar, and Jason Baldridge. 2019. PAWS-X: A cross-lingual adversarial dataset for paraphrase identification. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3687-3692, Hong Kong, China. Association for Computational Linguistics.  
+Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L. Griffiths, Yuan Cao, and Karthik Narasimhan. 2023. Tree of thoughts: Deliberate problem solving with large language models.  
+Fabio Massimo Zanzotto, Andrea Santilli, Leonardo Ranaldi, Dario Onorati, Pierfrancesco Tommasino, and Francesca Fallucchi. 2020. KERMIT: Complementing transformer architectures with encoders of explicit syntactic interpretations. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 256-267, Online. Association for Computational Linguistics.
+
+# A Prompt
+
+In this paper, we analyze our prompting approach, i.e., Cross-ToT, in different tasks. In Figure 1 we have shown the input-prompt for the MGSM (Cobbe et al., 2021). Here, we show the prompt framework for the other tasks:
+
+# XCOPA, XNLI and PAWS-X
+
+Simulate the collaboration of  $n$  person answering a question in their mother tongue:  $L_{1}$  and English. They all start Step1 from a separate thought process, step by step, each explaining their thought process. Following Step1, each expert refines and develops their thought process by comparing themselves with others. This process continues until a definitive answer to the question is obtained.
+
+# Basic Prompt
+
+Table 7: Our prompting approach for XCOPA, XNLI and PAWS-X. List of the Basic Prompt is in Table 11
+
+Furthermore, in Section 5.1, we proposed an experiment based on a prompt with only two languages as follows:
+
+# Binary Cross-ToT
+
+Simulate the collaboration of 2 mathematicians answering a question in their mother tongue:  $L_{1}$  and English. They all start Step1 from a separate thought process, step by step, each explaining their thought process. Following Step1, each expert refines and develops their thought process by comparing themselves with others. This process continues until a definitive answer to the question is obtained. Question: [Question in Language  $L_{1}$ ] Answer: [num].
+
+Table 8: Our prompting approach for experiment proposed in Section 5.1 regarding MGSM and binary trees
+
+# B Reasoning Chain
+
+# B.1 Chain-of-Thought Quality Scoring Implementation
+
+The ROSCOE framework (Golovneva et al., 2023) incorporates multiple chain-of-thought quality metrics, with the reasoning alignment vector  $\alpha$  that is
+
+$$
+r _ {\text {a l i g n}} (h \rightarrow s) = \left\{\alpha_ {1}, \alpha_ {2}, \dots , \alpha_ {N} \right\} \in [ 0, 1 ] ^ {N} \tag {1}
+$$
+
+from the  $N$ -step hypothesis  $h = \{h_i\}_{i=1}^N$  to the source input  $s$  of length  $T$ , where  $\alpha_i$  are defined as:  $r_{align}(h_i \to s) = \frac{1 + \max_{j=1}^{T} \cos(h_i, s_j)}{2}$
+
+Faithfulness score The Faithfulness  $(F)$  score is calculated based on the alignment between the hypothesis steps  $h$  and the source sentences  $s$ . It represents the average reasoning alignment score over the steps of reasoning:
+
+$$
+F = \frac {1}{N} \sum_ {i = 1} ^ {N} r _ {\text {a l i g n}} \left(h _ {i} \rightarrow s\right) \tag {2}
+$$
+
+The Faithfulness score serves as a measure to assess whether the model misconstrued the problem in the statement or if the reasoning chain is characterized by ambiguity, unimportance, or the misuse of information.
+
+Informativeness Informativeness-Step (Info-Step) measures the utilization of facts from the original text  $s$  in the reasoning steps  $h$ :
+
+$$
+I n f o _ {S t e p} = \frac {1}{2 T} \sum_ {t = 1} ^ {T} r _ {\text {a l i g n}} \left(s _ {t} \rightarrow h\right) + \frac {1}{2} F \tag {3}
+$$
+
+Info-Step assigns a higher score to reasoning steps that strongly align with the source, showing the capacity to which the generated hypothesis includes the information from the source. Conversely, a lower Info-Step score means reasoning steps unrelated to the source sentences or overlooking the provided information in the context.
+
+Informativeness Chain Like the Info-Step metric, the InformativenessChain (Info-Chain) metric estimates the degree of concordance between the hypothesis chain and the source. The calculation is as follows:
+
+$$
+I n f o _ {C h a i n} = \frac {1 + \cos (h , s)}{2} \tag {4}
+$$
+
+Missing Step The Missing Step (Miss-Step) metric is introduced to estimate any significant lacking steps, which examines the alignment between the reference reasoning text  $r = \{r_i\}^K$  and the hypothesis  $h$ . A miss-step is needed to meticulously assess each step in the reference and verify the existence of a similar step in the hypothesis. The metric is computed as:
+
+$$
+\text {M i s s - S t e p} = \min  _ {i = 1} ^ {K} (\mathrm {r} - \operatorname {a l i g n} (r _ {i}, h)). \tag {5}
+$$
+
+# C Other Results
+
+<table><tr><td># of shot-Cross-ToT</td><td>de</td><td>zh</td><td>fr</td><td>ru</td><td>sw</td><td>es</td><td>bn</td><td>ja</td><td>te</td><td>th</td><td>Avg</td></tr><tr><td>0-shot</td><td>86.5</td><td>84.2</td><td>83.9</td><td>83.2</td><td>74.3</td><td>84.4</td><td>78.7</td><td>79.8</td><td>68.7</td><td>74.6</td><td>79.8</td></tr><tr><td>1-shot</td><td>87.2</td><td>84.9</td><td>85.8</td><td>85.3</td><td>76.4</td><td>85.2</td><td>81.2</td><td>81.3</td><td>70.5</td><td>75.5</td><td>79.9</td></tr><tr><td>3-shot</td><td>88.4</td><td>85.7</td><td>87.2</td><td>87.5</td><td>77.3</td><td>87.3</td><td>82.3</td><td>81.5</td><td>70.3</td><td>76.9</td><td>83.4</td></tr></table>
+
+Table 9: Accuracies (%) on MGSM using zero-shot, one-shot and three-shot  
+
+<table><tr><td>Model</td><td>et</td><td>ht</td><td>id</td><td>it</td><td>qu</td><td>sw</td><td>ta</td><td>th</td><td>tr</td><td>vi</td><td>zh</td><td>Avg</td></tr><tr><td colspan="13">Bloomz-7B (Muennighoff et al., 2022)</td></tr><tr><td>En-CoT</td><td>21.8</td><td>24.2</td><td>50.6</td><td>41.6</td><td>41.4</td><td>48.6</td><td>53.8</td><td>38.4</td><td>37.6</td><td>47.0</td><td>64.2</td><td>42.7</td></tr><tr><td>CLP (Qin et al., 2023)</td><td>49.0</td><td>49.6</td><td>58.0</td><td>48.8</td><td>50.6</td><td>47.6</td><td>57.8</td><td>52.0</td><td>50.2</td><td>45.2</td><td>54.2</td><td>51.2</td></tr><tr><td>Cross-ToT</td><td>48.0</td><td>47.3</td><td>58.2</td><td>47.8</td><td>49.3</td><td>46.4</td><td>55.2</td><td>53.1</td><td>50.8</td><td>44.2</td><td>50.3</td><td>49.5</td></tr><tr><td colspan="13">llama-2-13B (Touvron et al., 2023)</td></tr><tr><td>En-CoT</td><td>39.6</td><td>32.5</td><td>58.4</td><td>55.8</td><td>47.2</td><td>34.6</td><td>47.4</td><td>33.2</td><td>43.0</td><td>59.6</td><td>50.4</td><td>45.6</td></tr><tr><td>CLP (Qin et al., 2023)</td><td>44.8</td><td>48.2</td><td>64.4</td><td>70.2</td><td>46.6</td><td>47.0</td><td>47.8</td><td>46.4</td><td>51.2</td><td>58.8</td><td>51.4</td><td>52.4</td></tr><tr><td>Cross-ToT</td><td>43.3</td><td>49.1</td><td>61.5</td><td>65.8</td><td>44.4</td><td>46.6</td><td>43.7</td><td>42.2</td><td>49.5</td><td>55.2</td><td>48.2</td><td>50.6</td></tr></table>
+
+# D Prompt Table
+
+Table 10: Comparison of smaller open-source models on XCOPA.  
+
+<table><tr><td>Benchmark</td><td>#Test</td><td>Basic Prompt</td></tr><tr><td>MGSM</td><td>250</td><td>Question: {problem}</td></tr><tr><td>XCOPA</td><td>200</td><td>Here is a premise: {premise}. What is the {question}? Help me pick the more plausible option: -choice1: {choice1}, -choice2: {choice2}</td></tr><tr><td>XNLI</td><td>200</td><td>{premise}. Based on the previous passage, is it true that {hypothesis}? Yes, No, or Maybe?</td></tr><tr><td>PAWS-X</td><td>200</td><td>Sentence 1: {sentence1} Sentence 2: {sentence2} Question: Does Sentence 1 paraphrase Sentence 2? Yes or No?</td></tr></table>
+
+# E Number of Languages
+
+Table 11: The basic prompt of each benchmark. #Test denotes the number of instances in the test set that we randomly selected due to the cost constraint excepted for MGSM.  
+
+<table><tr><td>Language</td><td>Percentage</td></tr><tr><td>English (en)</td><td>46.3%</td></tr><tr><td>Russian (ru)</td><td>6.0%</td></tr><tr><td>German (de)</td><td>5.4%</td></tr><tr><td>Chinese (zh)</td><td>5.3%</td></tr><tr><td>French (fr)</td><td>4.4%</td></tr><tr><td>Japanese (ja)</td><td>4.3%</td></tr><tr><td>Spanish (es)</td><td>4.2%</td></tr><tr><td>Other</td><td>23.1%</td></tr></table>
+
+Table 12: Language distribution of CommonCrawl (Common Crawl, 2021).

2024/A Tree-of-Thoughts to Broaden Multi-step Reasoning across Languages/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:14f808a7359e99c605c7502c4bb233f2df853089aaccfe6141ed72be2ab8df59
+size 594684